Release verified Image21-INT8 conversion

.gitattributes

@@ -33,3 +33,36 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/chinese_text-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/chinese_text-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/composition-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/composition-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/edit-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/edit-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/english_text-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/english_text-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/portrait-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/portrait-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/rgba-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/rgba-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/texture-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/bf16/texture-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/chinese_text-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/chinese_text-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/composition-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/composition-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/edit-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/edit-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/english_text-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/english_text-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/portrait-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/portrait-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/rgba-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/rgba-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/texture-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/int8/texture-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/young_woman/bf16/young_chinese_woman-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/young_woman/bf16/young_chinese_woman-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/young_woman/int8/young_chinese_woman-s123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/young_woman/int8/young_chinese_woman-s42.png filter=lfs diff=lfs merge=lfs -text
+processor/tokenizer.json filter=lfs diff=lfs merge=lfs -text

CHANGES.md

@@ -0,0 +1,9 @@
+# Modifications
+
+Modified by ixim / iximbox: eligible linear weights converted from Qwen-Image-2.1 to bitsandbytes LLM.int8 INT8. Built with Qwen. Non-commercial research/evaluation under the accompanying Qwen Research License.
+
+- Converted eligible transformer and text encoder linear layers to LLM.int8, threshold 6.0.
+- Retained sensitive projections, normalization, embeddings, vision model and VAE in floating point.
+- Re-serialized component weights, shard indexes and configs; no fine-tuning.
+- Weight-file headers and modified JSON files contain modification notices.
+- Exact quantized module names and dtype counts are in the component reports.

LICENSE

@@ -0,0 +1,55 @@
+Qwen RESEARCH LICENSE AGREEMENT
+
+Qwen RESEARCH LICENSE AGREEMENT Release Date: September 20, 2026
+
+By clicking to agree or by using or distributing any portion or element of the Qwen Materials, you will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
+
+1. Definitions
+    a. This Qwen RESEARCH LICENSE AGREEMENT (this "Agreement") shall mean the terms and conditions for use, reproduction, distribution and modification of the Materials as defined by this Agreement.
+    b. "We" (or "Us") shall mean Hangzhou Tongyi Laboratory Technology Co., Ltd.
+    c. "You" (or "Your") shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Materials for any purpose and in any field of use.
+    d. "Third Parties" shall mean individuals or legal entities that are not under common control with us or you.
+    e. "Qwen" shall mean the large language models, diffusion models, and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by us.
+    f. "Materials" shall mean, collectively, our proprietary Qwen and Documentation (and any portion thereof) made available under this Agreement.
+    g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
+    h. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+    i. "Non-Commercial" shall mean for research or evaluation purposes only.
+
+2. Grant of Rights
+    a. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under our intellectual property or other rights owned by us embodied in the Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials FOR NON-COMMERCIAL PURPOSES ONLY. 
+    b. You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us. If you wish to use the Materials commercially, you shall request a license from us at model-business@notice.qwencloud.com.
+
+3. Redistribution
+Subject to Section 2 (Grant of Rights), you may distribute copies or make the Materials, or derivative works thereof, available as part of a product or service that contains any of them, with or without modifications, and in Source or Object form, provided that you meet the following conditions:
+    a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
+    b. You shall cause any modified files to carry prominent notices stating that you changed the files;
+    c. You shall retain in all copies of the Materials that you distribute the following attribution notices within a "Notice" text file distributed as a part of such copies: "Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved."; and
+    d. You may add your own copyright statement to your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of your modifications, or for any such derivative works as a whole, provided your use, reproduction, and distribution of the work otherwise complies with the terms and conditions of this Agreement.
+
+4. Rules of use
+    a. The Materials may be subject to export controls or restrictions in China, the United States or other countries or regions. You shall comply with applicable laws and regulations in your use of the Materials.
+    b. If you use the Materials or any outputs or results therefrom to create, train, fine-tune, or improve an AI model that is distributed or made available, you shall prominently display “Built with Qwen” or “Improved using Qwen” in the related product documentation.
+    c. You shall not use "Qwen" as the primary name or identifier of any derivative works or products; reasonable descriptive use (e.g., "fine-tuned from Qwen Image") is permitted.
+
+5. Intellectual Property
+    a. We retain ownership of all intellectual property rights in and to the Materials and derivatives made by or for us. Conditioned upon compliance with the terms and conditions of this Agreement, with respect to any derivative works and modifications of the Materials that are made by you, you are and will be the owner of such derivative works and modifications.
+    b. No trademark license is granted to use the trade names, trademarks, service marks, or product names of us, except as required to fulfill notice requirements under this Agreement or as required for reasonable and customary use in describing and redistributing the Materials.
+    c. If you commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against us or any entity alleging that the Materials or any output therefrom, or any part of the foregoing, infringe any intellectual property or other right owned or licensable by you, then all licenses granted to you under this Agreement shall terminate as of the date such lawsuit or other proceeding is commenced or brought.
+
+6. Disclaimer of Warranty and Limitation of Liability
+    a. We are not obligated to support, update, provide training for, or develop any further version of the Qwen Materials or to grant any license thereto.
+    b. THE MATERIALS ARE PROVIDED "AS IS" WITHOUT ANY EXPRESS OR IMPLIED WARRANTY OF ANY KIND INCLUDING WARRANTIES OF MERCHANTABILITY, NONINFRINGEMENT, OR FITNESS FOR A PARTICULAR PURPOSE. WE MAKE NO WARRANTY AND ASSUME NO RESPONSIBILITY FOR THE SAFETY OR STABILITY OF THE MATERIALS AND ANY OUTPUT THEREFROM.
+    c. IN NO EVENT SHALL WE BE LIABLE TO YOU FOR ANY DAMAGES, INCLUDING, BUT NOT LIMITED TO ANY DIRECT, OR INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM YOUR USE OR INABILITY TO USE THE MATERIALS OR ANY OUTPUT OF IT, NO MATTER HOW IT’S CAUSED.
+    d. You will defend, indemnify and hold harmless us from and against any claim by any third party arising out of or related to your use or distribution of the Materials.
+
+7. Survival and Termination.
+    a. The term of this Agreement shall commence upon your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
+    b. We may terminate this Agreement if you breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, you must delete and cease use of the Materials. Sections 6 and 8 shall survive the termination of this Agreement.
+
+8. Governing Law and Jurisdiction.
+    a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
+    b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.
+
+9. Other Terms and Conditions.
+    a. Any arrangements, understandings, or agreements regarding the Material not stated herein are separate from and independent of the terms and conditions of this Agreement. You shall request a separate license from us, if you use the Materials in ways not expressly agreed to in this Agreement. 
+    b. We shall not be bound by any additional or different terms or conditions communicated by you unless expressly agreed.

MANIFEST.json

@@ -0,0 +1,532 @@
+[
+  {
+    "path": "CHANGES.md",
+    "size": 663,
+    "sha256": "c5e50b15f690fce48eaf9a837be7033b5ee8fdae89048b3cd3fef20df412e775"
+  },
+  {
+    "path": "LICENSE",
+    "size": 7831,
+    "sha256": "8dc973f024ff95966bea25866efa443fd16776dcb1001e681e3d467ea572b28d"
+  },
+  {
+    "path": "Notice",
+    "size": 530,
+    "sha256": "9b12c699d8e464d0eda682b18535568981893eaf66e462a0467f2c16b387970a"
+  },
+  {
+    "path": "PUBLISHING.md",
+    "size": 2118,
+    "sha256": "ed220160bb06dd4774f6e413eb6b077f8316033ae76dbf85c146e9c65cbac4e8"
+  },
+  {
+    "path": "README.md",
+    "size": 25079,
+    "sha256": "6250a13203f365d078fab9b8a338904dc81f2ca4100412c60b723b33c154f704"
+  },
+  {
+    "path": "REPRODUCE.md",
+    "size": 5400,
+    "sha256": "a5e66fe633b1d8e920c7192f7563dbdc72619d9e61c1ed5172b96ac49d790ffc"
+  },
+  {
+    "path": "benchmarks/cases.json",
+    "size": 1919,
+    "sha256": "2663ee3d489e1eb136e02b083963b8622146fc65ac301d6acc5a6aab21e63356"
+  },
+  {
+    "path": "benchmarks/young_woman.json",
+    "size": 516,
+    "sha256": "0d1d27b403e6c6685d652ef5553978fc9ac8af9b1d13c5c662449d0a8be74847"
+  },
+  {
+    "path": "cards/huggingface.md",
+    "size": 4370,
+    "sha256": "aef950ac712df59df5d679177aafebd9f8f0e7b772f2b2b6077b7aa8ce3a5fef"
+  },
+  {
+    "path": "cards/modelscope.md",
+    "size": 4139,
+    "sha256": "b4f5c38e7d2cbc2c3aeee9884f4136591b1083a1c8e4846f5b71a772cc119487"
+  },
+  {
+    "path": "conversion.json",
+    "size": 1425,
+    "sha256": "85900407b59658abab04a37e830089939ca20346435dea3ae3a09cb1e5f6887f"
+  },
+  {
+    "path": "evaluation/bf16/chinese_text-s123.png",
+    "size": 1194079,
+    "sha256": "eae74256d7ed1b3519d39c7d58961cf7a937224d19a3c3239ca7eba4178edf49"
+  },
+  {
+    "path": "evaluation/bf16/chinese_text-s42.png",
+    "size": 1303893,
+    "sha256": "7082cb71d99705bc42e27e2c206b13e3beb9e62cfd6304cfabe203b82641692c"
+  },
+  {
+    "path": "evaluation/bf16/composition-s123.png",
+    "size": 1220815,
+    "sha256": "bc11612f260acd8af0d834f3b28c752cefd0437d24c9a01dae0d46647bdf29ae"
+  },
+  {
+    "path": "evaluation/bf16/composition-s42.png",
+    "size": 1194343,
+    "sha256": "91e967685fbbd79ba13daece22a0f72c863cb518db88d5d2a6f9aecfc5f292b1"
+  },
+  {
+    "path": "evaluation/bf16/edit-s123.png",
+    "size": 1801031,
+    "sha256": "7d00c5bb7a2b31f05ea3fcf3f4a79dd93335770fe1b2d0ad772d5df954c9f91c"
+  },
+  {
+    "path": "evaluation/bf16/edit-s42.png",
+    "size": 2269130,
+    "sha256": "8713f6ef2ec0191d88225f2515671d6f51be43aa293bb1ca6e25b07a1b67deb9"
+  },
+  {
+    "path": "evaluation/bf16/english_text-s123.png",
+    "size": 905011,
+    "sha256": "cf5e76dee9f0e403ebdf24a64e1fcd604ee2b6c8e645a032fba2807acc34adf2"
+  },
+  {
+    "path": "evaluation/bf16/english_text-s42.png",
+    "size": 969130,
+    "sha256": "ba31b0c8a312e96408096566013c667b3a32bd7720749ca47120960b17b41ca6"
+  },
+  {
+    "path": "evaluation/bf16/environment.json",
+    "size": 5659,
+    "sha256": "9fe2336fb834a19c56b3280ad01d22bb3a716ddead12633398c8561eb6ffc981"
+  },
+  {
+    "path": "evaluation/bf16/portrait-s123.png",
+    "size": 1930022,
+    "sha256": "cff376e50c41c4aee2b9e4aabc502f4791774f6ad285d8a7673a61ebb1fa650d"
+  },
+  {
+    "path": "evaluation/bf16/portrait-s42.png",
+    "size": 1868618,
+    "sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa"
+  },
+  {
+    "path": "evaluation/bf16/records.jsonl",
+    "size": 11284,
+    "sha256": "4167b569c5970648ef04a7f903c12ba6bf658f0b6fb3931e79efdf98bbfddabb"
+  },
+  {
+    "path": "evaluation/bf16/rgba-s123.png",
+    "size": 926957,
+    "sha256": "6f1abbb1a64ac553fc576a0250d370ab5bb7e6d5e6372f42f67045a5bb0951a6"
+  },
+  {
+    "path": "evaluation/bf16/rgba-s42.png",
+    "size": 850131,
+    "sha256": "6f5dfea6ba02e41cbf20617b6e863e1adfa139d79fad2ca5c64dfc815f591d82"
+  },
+  {
+    "path": "evaluation/bf16/texture-s123.png",
+    "size": 1760350,
+    "sha256": "18c4403f388d0e03447473c03497141169f333fddeaeb40227163dc3b9d68d3e"
+  },
+  {
+    "path": "evaluation/bf16/texture-s42.png",
+    "size": 1797822,
+    "sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095"
+  },
+  {
+    "path": "evaluation/comparison.csv",
+    "size": 2510,
+    "sha256": "e441773cf19d5e6e7f8320f9da5e4ec2e2d84a991589c105424b240d1f7e34bb"
+  },
+  {
+    "path": "evaluation/comparison.html",
+    "size": 5966,
+    "sha256": "83bbd685fe09c581d1e675b5324681c9bc6944ef47687894095573576473b8db"
+  },
+  {
+    "path": "evaluation/hardware.json",
+    "size": 349,
+    "sha256": "3f18698ee838a786f10a3fdbcc31621995811797a30d7c803be8fc4ee8e5a6db"
+  },
+  {
+    "path": "evaluation/int8/chinese_text-s123.png",
+    "size": 1201749,
+    "sha256": "7ff8508a9454650920591c0aad34d1793501f609e853f2f8584c15eee5127479"
+  },
+  {
+    "path": "evaluation/int8/chinese_text-s42.png",
+    "size": 1315329,
+    "sha256": "dce2de1e3b8b3836c7584854c7cf340cf05bcb141fa8e3988264c62e1b20238c"
+  },
+  {
+    "path": "evaluation/int8/composition-s123.png",
+    "size": 1211968,
+    "sha256": "c030c007a9c3962533ffeb256d9dac619cc082a20a21ff6f9eb0c0ab7b7cadae"
+  },
+  {
+    "path": "evaluation/int8/composition-s42.png",
+    "size": 1222325,
+    "sha256": "5704252c99e7f9dbe9e512eb82bf4fd93650e26c1d6479412679e7868637bcdd"
+  },
+  {
+    "path": "evaluation/int8/edit-s123.png",
+    "size": 1799329,
+    "sha256": "d8023e34ba10a22b36f63d050c50d44f149cebc7fc8f7f612a47f27f2a4cb70b"
+  },
+  {
+    "path": "evaluation/int8/edit-s42.png",
+    "size": 2275894,
+    "sha256": "79725e243ccfd82b5387beffe6e1f37193ce9137b576b2ec20e8543547dd9d33"
+  },
+  {
+    "path": "evaluation/int8/english_text-s123.png",
+    "size": 922494,
+    "sha256": "b55551e7c17975ee90ad895e449379355886c3422d6dcdee9b0c3a4ce1c4632f"
+  },
+  {
+    "path": "evaluation/int8/english_text-s42.png",
+    "size": 960968,
+    "sha256": "98ea66acd842498e192bc69503e32f5f61a4e5e56887d0f86e9c936b0b0d976f"
+  },
+  {
+    "path": "evaluation/int8/environment.json",
+    "size": 5290,
+    "sha256": "44625db8d61251f534c3abf6e0a4cd3d00955f57986420d8673f3962d41a6d27"
+  },
+  {
+    "path": "evaluation/int8/portrait-s123.png",
+    "size": 1930928,
+    "sha256": "09391672b7aa15abc92a49184c94cd7dc09b570b2134c95f0a1524701630323c"
+  },
+  {
+    "path": "evaluation/int8/portrait-s42.png",
+    "size": 1865119,
+    "sha256": "e37187a2530d0f45378f17ca8fff47a7e174f1ce0c19ff89b9df8f4db1845acd"
+  },
+  {
+    "path": "evaluation/int8/records.jsonl",
+    "size": 11281,
+    "sha256": "c24ceada57729199c64f5d89ceae27f9b403b93fe777179d9b635428131033cd"
+  },
+  {
+    "path": "evaluation/int8/rgba-s123.png",
+    "size": 1032567,
+    "sha256": "4871aab8f9d404a479385053bdb4c18de7f4119884bb0e30997d7be721a69b45"
+  },
+  {
+    "path": "evaluation/int8/rgba-s42.png",
+    "size": 918813,
+    "sha256": "8624d495fe40f602a492b3366ccb0354d5684e4e084281503fdb435d0f4e265e"
+  },
+  {
+    "path": "evaluation/int8/texture-s123.png",
+    "size": 1758127,
+    "sha256": "82fc88d5ab3cd4201634c54567b327f200231321dd0207990ff4c893e78496c7"
+  },
+  {
+    "path": "evaluation/int8/texture-s42.png",
+    "size": 1784440,
+    "sha256": "37042d1659147a142e0562aa27623cb1d159d9d12e46af59395029f0293fd9c8"
+  },
+  {
+    "path": "evaluation/qualitative.md",
+    "size": 2365,
+    "sha256": "f5e112ed3eb9f64574e269489b0176a6d389b1e983e3b9d97c802d329bca7b4d"
+  },
+  {
+    "path": "evaluation/report.md",
+    "size": 2184,
+    "sha256": "84c4e4348c035e3852155ccb4a768441811c6c08ec8358b8f28c21ca190af5e8"
+  },
+  {
+    "path": "evaluation/summary.json",
+    "size": 365,
+    "sha256": "ab07b54644317bfc2caa1c0ad1b4d960e414752a37ad7e842d41685dcf523e38"
+  },
+  {
+    "path": "evaluation/young_woman/bf16/environment.json",
+    "size": 5480,
+    "sha256": "f22b1f0a4e9eb98d6e402e0d08b93afc595f6391378071c59a827c172b131c84"
+  },
+  {
+    "path": "evaluation/young_woman/bf16/records.jsonl",
+    "size": 2067,
+    "sha256": "881dd5592fc198ede61416a9f7a6dcf9dc697ae5e678ffb166e04a52192f831c"
+  },
+  {
+    "path": "evaluation/young_woman/bf16/young_chinese_woman-s123.png",
+    "size": 1555909,
+    "sha256": "a2f28b40c897335595e496d23e490278c53b2f916597a4ef0fbb95a9f2700096"
+  },
+  {
+    "path": "evaluation/young_woman/bf16/young_chinese_woman-s42.png",
+    "size": 1568091,
+    "sha256": "00716ad4721f585be4b0bf0daf2413f4ab92b795f1eb4d08b4f1ecc604b075b8"
+  },
+  {
+    "path": "evaluation/young_woman/comparison.csv",
+    "size": 513,
+    "sha256": "c38eef422e781fb0b3cddeee017027140c62eaa2abfa2e6948a32e184a933944"
+  },
+  {
+    "path": "evaluation/young_woman/comparison.html",
+    "size": 1726,
+    "sha256": "3586888ee7622d2226661184cfdf5493200c3853c65e7c0890f4365fccf6fc26"
+  },
+  {
+    "path": "evaluation/young_woman/int8/environment.json",
+    "size": 5289,
+    "sha256": "703186d40488bba0c1a058bf0134b2436ebba1b0929679e1b71cb78c4c292722"
+  },
+  {
+    "path": "evaluation/young_woman/int8/records.jsonl",
+    "size": 2064,
+    "sha256": "b3bf88a2f62ac9262c94730e92fb2d3816f4a13bab1c9c6716c4b991f0bff3b7"
+  },
+  {
+    "path": "evaluation/young_woman/int8/young_chinese_woman-s123.png",
+    "size": 1560080,
+    "sha256": "350e800c857e0c6e0836cd7fb222cd9b60c10a34f29f15b1f8ffa71e9f986f64"
+  },
+  {
+    "path": "evaluation/young_woman/int8/young_chinese_woman-s42.png",
+    "size": 1563866,
+    "sha256": "5ffeb0f35bf0115ba2dd7ea9a98def2a97f280c67c27a24c9d1110e6501256a6"
+  },
+  {
+    "path": "evaluation/young_woman/qualitative.md",
+    "size": 2185,
+    "sha256": "98ad6ce5f287a201d375ebfcf0c6fd8b3a4c254d44ae9f565446b661f6eb324a"
+  },
+  {
+    "path": "evaluation/young_woman/report.md",
+    "size": 1300,
+    "sha256": "b647a04ecf6af54fc6cdcfe1ca34383f34b513be01f5b2c3048db6daa2142bde"
+  },
+  {
+    "path": "evaluation/young_woman/summary.json",
+    "size": 365,
+    "sha256": "f915a59019a6ebdfe8761dc7f9bc30e7348ecf1ef1e83e899f808b9d6af7e7ed"
+  },
+  {
+    "path": "model_index.json",
+    "size": 447,
+    "sha256": "cf1ecd104ea090855d60d8cec0895c1e9d6ee41b2f87e231b678beecd1cf7809"
+  },
+  {
+    "path": "processor/added_tokens.json",
+    "size": 707,
+    "sha256": "c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680"
+  },
+  {
+    "path": "processor/chat_template.jinja",
+    "size": 5292,
+    "sha256": "3636d0f0bd6bef02654cdffdc447b79cb2cef8ab02cc75267345946291a489e4"
+  },
+  {
+    "path": "processor/merges.txt",
+    "size": 1671853,
+    "sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
+  },
+  {
+    "path": "processor/preprocessor_config.json",
+    "size": 782,
+    "sha256": "93585062a80db5e8ca038efc7726a3e6411d9db948472d81d63c6303993be8c5"
+  },
+  {
+    "path": "processor/special_tokens_map.json",
+    "size": 613,
+    "sha256": "76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd"
+  },
+  {
+    "path": "processor/tokenizer.json",
+    "size": 11422654,
+    "sha256": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4"
+  },
+  {
+    "path": "processor/tokenizer_config.json",
+    "size": 5445,
+    "sha256": "81ec7bb9530159b326c0bef1d0b6c33d392090524014ea3f0123a3c1eb9c2af5"
+  },
+  {
+    "path": "processor/video_preprocessor_config.json",
+    "size": 817,
+    "sha256": "59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1"
+  },
+  {
+    "path": "processor/vocab.json",
+    "size": 2776833,
+    "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
+  },
+  {
+    "path": "requirements.txt",
+    "size": 358,
+    "sha256": "c9fa98a3948380f630abb7a84fcfc4e2dac020b0017a5e2cae672ebca3781f6c"
+  },
+  {
+    "path": "scheduler/scheduler_config.json",
+    "size": 485,
+    "sha256": "5895f3a167c14a967fe9ac70c64924ae5acc79799e0679fd12907e594a713cd1"
+  },
+  {
+    "path": "scripts/__init__.py",
+    "size": 364,
+    "sha256": "bf9f0329f310b94e93f498ffb6e921604dba02340d7b77168f5da0d8384f5762"
+  },
+  {
+    "path": "scripts/benchmark.py",
+    "size": 5726,
+    "sha256": "4e1bc7bef15e04b436fd5b32a21c33954319ae670de32c5eca33ede742c8dd01"
+  },
+  {
+    "path": "scripts/download.py",
+    "size": 2455,
+    "sha256": "0148190b086caa0e016a4a55cf4f6081ba1bf3721631342bdaecfdfab4f70565"
+  },
+  {
+    "path": "scripts/infer.py",
+    "size": 1829,
+    "sha256": "1f73851dd4c8bed57d0444ea195ef5008f3827f5f6974cae0c65fe710a3e2984"
+  },
+  {
+    "path": "scripts/integrity.py",
+    "size": 2136,
+    "sha256": "eecb338ebaed60a32e7a8ee8b9a6097a570f73af8da10bdfe67fe96b764f5d71"
+  },
+  {
+    "path": "scripts/model_card.py",
+    "size": 13708,
+    "sha256": "93839adc07c6b4400ad493f3400c7d0bf03e3673f83a25820e80c7c6fba731cd"
+  },
+  {
+    "path": "scripts/provenance.py",
+    "size": 4495,
+    "sha256": "9b0c2957ecd1260bee26716922d746244caa857c2183ae31288e1d2ad5e3ab26"
+  },
+  {
+    "path": "scripts/quantize.py",
+    "size": 4301,
+    "sha256": "1d12d3acd9c8e4f41cdd83f4be54b56228e1721ddf45bcee2732ff68e9259488"
+  },
+  {
+    "path": "scripts/release.py",
+    "size": 12517,
+    "sha256": "5dbd00f5d29be7f74c171352fedde6a4c2d807b359277b2b3bff401c0f5ddf0f"
+  },
+  {
+    "path": "scripts/report.py",
+    "size": 8179,
+    "sha256": "f33f8352a0b6561b8089b04d280f4082feb55de0c89250a0b02fc35639c94243"
+  },
+  {
+    "path": "scripts/upload.py",
+    "size": 3674,
+    "sha256": "b46f2c8595e7aff91e1565a1f068ab9c32c85f9daab8cb5383f9e5ba3024fc4a"
+  },
+  {
+    "path": "tests/__init__.py",
+    "size": 0,
+    "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+  },
+  {
+    "path": "tests/gpu_roundtrip.py",
+    "size": 1055,
+    "sha256": "7077bdac70c0f17882834023bc98c2c9e2eaa0e30c0fd8012e08052cd29df45b"
+  },
+  {
+    "path": "tests/test_integrity.py",
+    "size": 1739,
+    "sha256": "1042a18ea03c2b76a84584a51cd9f0a7416321ea999dbc9f1ec00c38e28e815b"
+  },
+  {
+    "path": "tests/test_notices.py",
+    "size": 1094,
+    "sha256": "0f850d1bdddc4c1ee431dae56961b439a566d10787d6968c6a17e1364edff290"
+  },
+  {
+    "path": "tests/test_provenance.py",
+    "size": 2180,
+    "sha256": "47a448161b9a2834610f4e04a094b0aa0af268b628b11386dae65591d5b01b0a"
+  },
+  {
+    "path": "tests/test_upload_auth.py",
+    "size": 858,
+    "sha256": "722994cc6aab6673498266d6414650a88837c331b261693f2c0eb98499cb3552"
+  },
+  {
+    "path": "text_encoder-quantization.json",
+    "size": 19004,
+    "sha256": "602398ec933e151e3e71b5f3208500b376855b40d072f73340490e2a6e64f32b"
+  },
+  {
+    "path": "text_encoder/config.json",
+    "size": 2487,
+    "sha256": "da4f7a17ffbcaef0d09f9c31cab01d77e2c25905676818279254b5a8bb371150"
+  },
+  {
+    "path": "text_encoder/generation_config.json",
+    "size": 467,
+    "sha256": "7dc6dd5e325f76cd2f75d428d41049e1f8ac0c4775d0800c0885b88bfe3f4c7f"
+  },
+  {
+    "path": "text_encoder/model-00001-of-00003.safetensors",
+    "size": 4999822155,
+    "sha256": "9b8ae3bd2a1e4884abc48c659212427bbbae06fb1bc17e7173583704fc0f6bdd"
+  },
+  {
+    "path": "text_encoder/model-00002-of-00003.safetensors",
+    "size": 4996432636,
+    "sha256": "d28d1a8b3994a43582acc5e2da15ff38b632fae1466f41bc7710d27ec170aa9e"
+  },
+  {
+    "path": "text_encoder/model-00003-of-00003.safetensors",
+    "size": 27729689,
+    "sha256": "5896ffa9e2b2a52f9d496755fde050306f21c449bd66c2526b40979033c6e244"
+  },
+  {
+    "path": "text_encoder/model.safetensors.index.json",
+    "size": 139297,
+    "sha256": "d3fc335d23047182f40d458efec86a4b187a766e0d1639426b6a9b7ab5b1ab8a"
+  },
+  {
+    "path": "transformer-quantization.json",
+    "size": 9853,
+    "sha256": "d735fe2c5f2546f20d943cc0789a5f57950d945343871a8cade62ea9234996a7"
+  },
+  {
+    "path": "transformer/config.json",
+    "size": 1297,
+    "sha256": "1a941b5389b5985cd16314641d708a1333a65748bcd3f74b1b8b1cd74b828202"
+  },
+  {
+    "path": "transformer/diffusion_pytorch_model-00001-of-00002.safetensors",
+    "size": 4989447305,
+    "sha256": "0d77daaa3129860f026567e0178fec4ed776f3cf2463da76a0823b9c381c3468"
+  },
+  {
+    "path": "transformer/diffusion_pytorch_model-00002-of-00002.safetensors",
+    "size": 2267331623,
+    "sha256": "5e1b8ee3917833687adb662830438dc612ace2cf4bff7b8898515184a888fde1"
+  },
+  {
+    "path": "transformer/diffusion_pytorch_model.safetensors.index.json",
+    "size": 77984,
+    "sha256": "96963ae89098cbfdb83fb09bfe0b91e34d8dfe812c65da5e1fbe452878a79ebe"
+  },
+  {
+    "path": "upstream-verification.json",
+    "size": 1439,
+    "sha256": "2878036fc34cd72df293bd8432518a7a9ffbf9fe98d9ba13e2e1ed3a32f878e5"
+  },
+  {
+    "path": "vae/config.json",
+    "size": 2079,
+    "sha256": "9785d527b278cb8b210e9f48a8028d92a4190968ce7e24c8fc85d9d1b82f6ba2"
+  },
+  {
+    "path": "vae/diffusion_pytorch_model.safetensors",
+    "size": 1350989512,
+    "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+  }
+]

Notice

@@ -0,0 +1,6 @@
+Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
+
+Built with Qwen
+Independent derivative: Image21-INT8, by ixim / iximbox.
+Modified files: transformer and text_encoder weight shards, shard indexes and config.json files. Eligible linear weights converted to bitsandbytes INT8; floating-point exceptions are documented in component quantization reports.
+VAE, scheduler and processor retained from the pinned upstream snapshot.

PUBLISHING.md

@@ -0,0 +1,48 @@
+# 发布步骤
+
+以下命令可发布或续传已验证的本地发布包；平台上的实际提交是发布完成的依据。
+
+- Hugging Face：`release/huggingface`，英文 model card，目标 `ixim/Image21-INT8`。
+- ModelScope：`release/modelscope`，中文 model card，目标 `iximbox/Image21-INT8`。
+- ModelScope SDK 登录与网页登录分开；上传需要有效 API cookies 或 SDK Token。
+- Hugging Face 已核验项目缓存中的登录账号为 `ixim`。
+
+请在 **D:/Git/Qwen-Image-2.1** 打开 PowerShell。以下操作不需要把 Token 发给聊天。
+
+## Hugging Face
+
+```powershell
+$env:HF_HOME = "$PWD/.cache/huggingface"
+.venv/Scripts/hf.exe auth login
+.venv/Scripts/python.exe -m scripts.upload huggingface
+```
+
+使用 `ixim` 账号的写权限 Token。脚本会检查实际登录账号、校验所有发布文件，
+创建公开模型仓库并上传。完成后查看：
+https://huggingface.co/ixim/Image21-INT8
+
+本机已完成登录时，无需重复 `auth login`；请保留项目 `HF_HOME` 设置后执行上传命令。
+
+## ModelScope
+
+```powershell
+.venv/Scripts/python.exe -c "from getpass import getpass; from modelscope.hub.api import HubApi; HubApi().login(getpass('ModelScope SDK token: '))"
+.venv/Scripts/python.exe -m scripts.upload modelscope
+```
+
+使用 `iximbox` 账号新取得的 SDK Token（输入不回显）。Git access token 不能替代
+SDK Token。登录后脚本使用 API cookies，检查账号，再创建/上传目标仓库。完成后查看：
+https://modelscope.cn/models/iximbox/Image21-INT8
+
+## 上传前仅校验
+
+```powershell
+.venv/Scripts/python.exe -m scripts.upload huggingface --dry-run
+.venv/Scripts/python.exe -m scripts.upload modelscope --dry-run
+```
+
+不要上传整个工作区；源 BF16、缓存和登录信息不属于发布包。中断上传后可再次执行
+对应上传命令。上传脚本不会删除远程文件。
+
+原模型使用 Qwen Research License，发布包已保留协议、归属与修改声明。
+衍生模型主名称为 Image21-INT8，仅用于非商业研究与评测。

README.md

@@ -0,0 +1,443 @@
+---
+license: other
+license_name: qwen-research
+license_link: LICENSE
+base_model: Qwen/Qwen-Image-2.1
+base_model_relation: quantized
+library_name: diffusers
+pipeline_tag: text-to-image
+language:
+  - en
+  - zh
+tags:
+  - diffusers
+  - bitsandbytes
+  - int8
+  - image-generation
+  - image-editing
+  - rgba
+---
+
+# Image21-INT8
+
+**Built with Qwen.** An independent bitsandbytes INT8 derivative of
+[Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1), prepared by **ixim**.
+This is a community conversion, not an official Qwen release.
+
+## Quantization
+
+Base revision: `b3179ad355be050328e483a9dfdd9e60cd62adfa`.
+Eligible linear layers in the generation transformer and the Qwen3-VL text encoder
+use **bitsandbytes LLM.int8**, with outlier threshold **6.0**. This is mixed-precision
+inference: outlier handling uses floating point; the VAE, normalization, embeddings,
+vision encoder, output head, and excluded conditioning/projection layers remain
+floating point. See the two component quantization JSON files for exact coverage.
+No fine-tuning or calibration dataset was used. This is not a pure W8A8 pipeline,
+FP8 conversion, or GGUF checkpoint. Lower storage does not imply faster inference.
+
+## Installation and loading
+
+Use Python 3.13 and a CUDA-capable PyTorch installation. The tested runtime uses
+PyTorch 2.10.0+cu128 on an RTX 5090. Install the accompanying `requirements.txt`;
+it pins Diffusers to the official commit used for this conversion.
+
+```bash
+pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
+pip install -r requirements.txt
+```
+
+```python
+import torch
+from diffusers import QwenImage21Pipeline
+
+pipe = QwenImage21Pipeline.from_pretrained(
+    "ixim/Image21-INT8", dtype=torch.bfloat16,
+)
+pipe.enable_model_cpu_offload()
+image = pipe(
+    prompt='A neon sign reading "CREATE WITH LIGHT", rainy night',
+    width=1024, height=1024, num_inference_steps=40,
+    true_cfg_scale=1.0, use_kv_cache=True,
+    generator=torch.Generator("cpu").manual_seed(42),
+).images[0]
+image.save("output.png")
+```
+
+The saved quantization configuration is detected automatically. Do not re-quantize
+these weights at load time. For editing, pass `image=PIL.Image.open("input.png")`.
+For transparency, use the upstream RGBA prompt format and save PNG files.
+
+## Local evaluation
+
+14 paired outputs across 7 cases at 1024×1024 on an RTX 5090.
+
+| Metric | BF16 | INT8 |
+|---|---:|---:|
+| Weight files (decimal GB) | 33.116 | 18.632 |
+| Mean call latency (s) | 26.79 | 20.29 |
+| Maximum CUDA allocated memory (GiB) | 19.08 | 23.08 |
+
+**Observed trade-off:** weights are smaller and average latency is lower in this
+run, but peak allocated GPU memory is **higher** with the tested INT8/offload stack.
+Do not use this checkpoint as evidence of lower-VRAM operation.
+
+**Editing limitation:** both BF16 and INT8 sweater-edit examples show substantial
+oversharpening and contrast changes. Faithful preservation of the original image
+is not established. Both precisions also add unwanted footer text in the inspected
+Chinese poster. See [qualitative observations](evaluation/qualitative.md).
+
+The comparison uses identical prompts, CPU random generators, seeds, resolutions,
+40 denoising steps, CFG=1, KV cache enabled, and model CPU offload. Warmup is excluded.
+Peak allocated/reserved memory are PyTorch CUDA statistics, not total board memory.
+Timing includes prompt encoding, denoising, decoding and CPU transfers; it excludes
+model loading and image-file writing. Paired pixel differences measure drift, not
+semantic quality. This small suite is not a standardized quality ranking.
+BF16 was run first, followed by INT8 on the same Windows desktop GPU; repeated
+timing trials and confidence intervals were not collected.
+
+See [evaluation/report.md](evaluation/report.md), raw JSONL records, environment
+records and paired PNG outputs. The edit pair uses the exact same source image.
+Higher resolutions, extensive multi-reference editing, other GPUs, other attention
+backends, and ComfyUI compatibility are not established by this evaluation.
+
+### Evaluation scope and runtime
+
+The original suite has 14 pairs; the separate portrait supplement has 2 pairs. All 16 pairs / 32 original PNG images are shown below. Supplementary measurements are kept separate from the original summary average.
+
+All previews link to their full-resolution original PNG. Click an image to open it; no crop, retouching or regenerated substitute is used.
+
+| Setting | Recorded value |
+|---|---|
+| GPU | NVIDIA GeForce RTX 5090 |
+| Driver / board memory | 616.92 / 32607 MiB |
+| System RAM | 255.5 GiB |
+| OS | Windows-11-10.0.26200-SP0 |
+| Python / CUDA | 3.13.5 / 12.8 |
+| Generation | 1024×1024; 40 steps; CFG 1.0; KV cache enabled |
+| RNG / seeds | CPU torch.Generator; 42, 123 |
+| Offload / attention | model CPU offload / default SDPA |
+| Warmup / order | One excluded full-settings warmup per precision per suite; BF16 then INT8 |
+
+| Package | Version |
+|---|---|
+| torch | 2.10.0+cu128 |
+| torchvision | 0.25.0+cu128 |
+| diffusers | 0.41.0.dev0 |
+| transformers | 5.17.0 |
+| accelerate | 1.15.0 |
+| bitsandbytes | 0.50.2 |
+| safetensors | 0.8.0 |
+| huggingface-hub | 1.32.0 |
+| tokenizers | 0.23.2 |
+| numpy | 2.2.6 |
+| pillow | 12.2.0 |
+| psutil | 7.0.0 |
+
+### Supplementary portrait summary
+
+| Metric | BF16 | INT8 |
+|---|---:|---:|
+| Mean call latency (s) | 28.51 | 20.16 |
+| Maximum CUDA allocated (GiB) | 16.44 | 22.86 |
+
+### Per-image timing and memory
+
+| Case | Seed | BF16 s | INT8 s | BF16 allocated GiB | INT8 allocated GiB |
+|---|---:|---:|---:|---:|---:|
+| Older adult portrait | 42 | 29.77 | 19.65 | 16.40 | 22.86 |
+| Older adult portrait | 123 | 26.55 | 19.61 | 16.40 | 22.86 |
+| English lettering | 42 | 25.95 | 19.42 | 16.41 | 22.86 |
+| English lettering | 123 | 25.88 | 19.40 | 16.41 | 22.86 |
+| Chinese lettering | 42 | 25.47 | 19.54 | 16.41 | 22.86 |
+| Chinese lettering | 123 | 25.85 | 19.86 | 16.41 | 22.86 |
+| Counting and spatial arrangement | 42 | 25.38 | 19.13 | 16.41 | 22.86 |
+| Counting and spatial arrangement | 123 | 25.35 | 19.30 | 16.41 | 22.86 |
+| Feather and branch texture | 42 | 25.85 | 19.22 | 16.40 | 22.86 |
+| Feather and branch texture | 123 | 25.76 | 19.57 | 16.40 | 22.86 |
+| RGBA transparency | 42 | 26.04 | 19.39 | 16.40 | 22.86 |
+| RGBA transparency | 123 | 25.07 | 19.29 | 16.40 | 22.86 |
+| Image editing: sweater recoloring | 42 | 31.16 | 25.40 | 19.08 | 23.08 |
+| Image editing: sweater recoloring | 123 | 31.03 | 25.24 | 19.08 | 23.08 |
+| Young Chinese adult woman portrait (supplement) | 42 | 29.50 | 20.30 | 16.44 | 22.86 |
+| Young Chinese adult woman portrait (supplement) | 123 | 27.53 | 20.03 | 16.44 | 22.86 |
+
+### Pixel drift (not a semantic quality score)
+
+RGB MAE and PSNR use a white composite so invisible RGB values in transparent pixels do not dominate the comparison. Alpha MAE is measured separately.
+
+| Case | Seed | RGB MAE | RGB PSNR dB | Alpha MAE |
+|---|---:|---:|---:|---:|
+| Older adult portrait | 42 | 0.01341 | 31.70 | 0.000046 |
+| Older adult portrait | 123 | 0.01630 | 30.55 | 0.000180 |
+| English lettering | 42 | 0.02626 | 19.45 | 0.000008 |
+| English lettering | 123 | 0.01180 | 25.83 | 0.000010 |
+| Chinese lettering | 42 | 0.05972 | 15.60 | 0.000108 |
+| Chinese lettering | 123 | 0.02460 | 20.44 | 0.000068 |
+| Counting and spatial arrangement | 42 | 0.02268 | 23.67 | 0.000022 |
+| Counting and spatial arrangement | 123 | 0.01959 | 22.87 | 0.000035 |
+| Feather and branch texture | 42 | 0.06978 | 18.02 | 0.000191 |
+| Feather and branch texture | 123 | 0.01684 | 27.93 | 0.000059 |
+| RGBA transparency | 42 | 0.01829 | 20.89 | 0.011088 |
+| RGBA transparency | 123 | 0.04647 | 17.05 | 0.029543 |
+| Image editing: sweater recoloring | 42 | 0.00971 | 36.48 | 0.000111 |
+| Image editing: sweater recoloring | 123 | 0.00391 | 43.27 | 0.000026 |
+| Young Chinese adult woman portrait (supplement) | 42 | 0.01002 | 31.42 | 0.000121 |
+| Young Chinese adult woman portrait (supplement) | 123 | 0.00631 | 35.94 | 0.000093 |
+
+### Additional memory telemetry
+
+Reserved memory is the PyTorch CUDA allocator peak. CPU RSS is the process snapshot after inference, not a peak.
+
+| Case | Seed | BF16 reserved GiB | INT8 reserved GiB | BF16 RSS GiB | INT8 RSS GiB |
+|---|---:|---:|---:|---:|---:|
+| Older adult portrait | 42 | 16.58 | 22.91 | 39.33 | 23.17 |
+| Older adult portrait | 123 | 16.58 | 22.91 | 42.59 | 24.06 |
+| English lettering | 42 | 16.58 | 22.91 | 43.68 | 24.25 |
+| English lettering | 123 | 16.58 | 22.91 | 44.81 | 24.37 |
+| Chinese lettering | 42 | 16.59 | 22.91 | 45.28 | 24.50 |
+| Chinese lettering | 123 | 16.59 | 22.91 | 45.83 | 24.48 |
+| Counting and spatial arrangement | 42 | 16.59 | 22.91 | 45.99 | 24.46 |
+| Counting and spatial arrangement | 123 | 16.59 | 22.91 | 46.37 | 25.91 |
+| Feather and branch texture | 42 | 16.58 | 22.91 | 46.50 | 26.13 |
+| Feather and branch texture | 123 | 16.58 | 22.91 | 46.85 | 26.12 |
+| RGBA transparency | 42 | 16.58 | 22.91 | 47.62 | 26.14 |
+| RGBA transparency | 123 | 16.58 | 22.91 | 47.48 | 26.15 |
+| Image editing: sweater recoloring | 42 | 20.03 | 24.62 | 48.08 | 27.16 |
+| Image editing: sweater recoloring | 123 | 20.03 | 24.62 | 48.49 | 27.13 |
+| Young Chinese adult woman portrait (supplement) | 42 | 16.62 | 22.93 | 39.59 | 23.32 |
+| Young Chinese adult woman portrait (supplement) | 123 | 16.62 | 22.93 | 42.12 | 23.91 |
+
+### All paired samples, prompts and observations
+
+Left: BF16; right: saved and reloaded INT8. Same prompt and seed within each pair. The original seven-case observations primarily inspect seed 42; both supplementary seeds were inspected. This is not a blinded preference study or an OCR/CLIP/FID score.
+
+#### Young Chinese adult woman portrait (supplement)
+
+Exact measured prompt:
+
+```text
+一张写实摄影风格的中国年轻成年女性肖像，人物明确为25岁成年人。黑色及肩直发，穿着米白色圆领针织衫，胸部以上构图，正面看向镜头，神情自然放松，轻微微笑。窗边柔和的自然日光，浅灰色简洁背景，真实自然的皮肤纹理和少量雀斑，淡妆，不过度磨皮。85mm人像摄影，浅景深，色彩自然，画面中没有文字。
+```
+
+Both seeds retain the broad framing, clothing and window lighting. Expression, facial contours, skin texture and individual hair strands change. No obvious large facial distortion was seen in these four outputs; this is not an identity-preservation score.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Young Chinese adult woman portrait (supplement) · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/bf16/young_chinese_woman-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/bf16/young_chinese_woman-s42.png) | [![Young Chinese adult woman portrait (supplement) · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/int8/young_chinese_woman-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/int8/young_chinese_woman-s42.png) |
+
+SHA256 prefixes: BF16 `00716ad4721f` / INT8 `5ffeb0f35bf0`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Young Chinese adult woman portrait (supplement) · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/bf16/young_chinese_woman-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/bf16/young_chinese_woman-s123.png) | [![Young Chinese adult woman portrait (supplement) · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/int8/young_chinese_woman-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/young_woman/int8/young_chinese_woman-s123.png) |
+
+SHA256 prefixes: BF16 `a2f28b40c897` / INT8 `350e800c857e`.
+
+#### Older adult portrait
+
+Exact measured prompt:
+
+```text
+A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.
+```
+
+Both depict the requested older woman, blue sweater and window. Face shape, wrinkles, framing and fabric details shift.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Older adult portrait · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/portrait-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/portrait-s42.png) | [![Older adult portrait · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/portrait-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/portrait-s42.png) |
+
+SHA256 prefixes: BF16 `01e37e178fb0` / INT8 `e37187a2530d`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Older adult portrait · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/portrait-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/portrait-s123.png) | [![Older adult portrait · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/portrait-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/portrait-s123.png) |
+
+SHA256 prefixes: BF16 `cff376e50c41` / INT8 `09391672b7aa`.
+
+#### English lettering
+
+Exact measured prompt:
+
+```text
+A clean editorial poster with a deep blue background. Large exact headline at the top: "CREATE WITH LIGHT". Smaller exact subtitle: "September 2026". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.
+```
+
+Both inspected outputs visibly render CREATE WITH LIGHT and September 2026. Layout and desk objects differ.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![English lettering · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/english_text-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/english_text-s42.png) | [![English lettering · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/english_text-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/english_text-s42.png) |
+
+SHA256 prefixes: BF16 `ba31b0c8a312` / INT8 `98ea66acd842`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![English lettering · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/english_text-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/english_text-s123.png) | [![English lettering · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/english_text-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/english_text-s123.png) |
+
+SHA256 prefixes: BF16 `cf5e76dee9f0` / INT8 `b55551e7c179`.
+
+#### Chinese lettering
+
+Exact measured prompt:
+
+```text
+设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。
+```
+
+Both inspected outputs render the requested main title and subtitle, but also add unwanted, partly nonsensical footer text despite the no-extra-text instruction.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Chinese lettering · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/chinese_text-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/chinese_text-s42.png) | [![Chinese lettering · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/chinese_text-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/chinese_text-s42.png) |
+
+SHA256 prefixes: BF16 `7082cb71d997` / INT8 `dce2de1e3b8b`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Chinese lettering · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/chinese_text-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/chinese_text-s123.png) | [![Chinese lettering · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/chinese_text-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/chinese_text-s123.png) |
+
+SHA256 prefixes: BF16 `eae74256d7ed` / INT8 `7ff8508a9454`.
+
+#### Counting and spatial arrangement
+
+Exact measured prompt:
+
+```text
+A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.
+```
+
+Both show three cups in red/blue/yellow order and one green apple in front of the blue cup. Object sizes and spacing differ.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Counting and spatial arrangement · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/composition-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/composition-s42.png) | [![Counting and spatial arrangement · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/composition-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/composition-s42.png) |
+
+SHA256 prefixes: BF16 `91e967685fbb` / INT8 `5704252c99e7`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Counting and spatial arrangement · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/composition-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/composition-s123.png) | [![Counting and spatial arrangement · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/composition-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/composition-s123.png) |
+
+SHA256 prefixes: BF16 `bc11612f260a` / INT8 `c030c007a9c3`.
+
+#### Feather and branch texture
+
+Exact measured prompt:
+
+```text
+Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.
+```
+
+Both depict a plausible kingfisher; head angle, branch orientation and feather texture differ. Pixel drift alone does not establish worse semantic quality.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Feather and branch texture · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/texture-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/texture-s42.png) | [![Feather and branch texture · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/texture-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/texture-s42.png) |
+
+SHA256 prefixes: BF16 `40a0f73bb407` / INT8 `37042d165914`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Feather and branch texture · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/texture-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/texture-s123.png) | [![Feather and branch texture · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/texture-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/texture-s123.png) |
+
+SHA256 prefixes: BF16 `18c4403f388d` / INT8 `82fc88d5ab3c`.
+
+#### RGBA transparency
+
+Exact measured prompt:
+
+```text
+This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.
+```
+
+Both have a real varying alpha channel and a white sticker outline. Sampled background alpha is 0–1/255 and the subject is opaque. Page background colors are not part of the PNG; click through to inspect the original RGBA file.
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![RGBA transparency · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/rgba-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/rgba-s42.png) | [![RGBA transparency · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/rgba-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/rgba-s42.png) |
+
+SHA256 prefixes: BF16 `6f5dfea6ba02` / INT8 `8624d495fe40`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![RGBA transparency · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/rgba-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/rgba-s123.png) | [![RGBA transparency · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/rgba-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/rgba-s123.png) |
+
+SHA256 prefixes: BF16 `6f1abbb1a64a` / INT8 `4871aab8f9d4`.
+
+#### Image editing: sweater recoloring
+
+Exact measured prompt:
+
+```text
+Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.
+```
+
+Both shift the sweater toward red/burgundy, but both also show conspicuous oversharpening and contrast/texture changes across the image. Faithful preservation of lighting and skin texture is not achieved. The cause is not identified or attributed specifically to INT8.
+
+Both variants use [this identical BF16 portrait input](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/portrait-s42.png).
+
+**Seed 42**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Image editing: sweater recoloring · BF16 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/edit-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/edit-s42.png) | [![Image editing: sweater recoloring · INT8 · seed 42](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/edit-s42.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/edit-s42.png) |
+
+SHA256 prefixes: BF16 `8713f6ef2ec0` / INT8 `79725e243ccf`.
+
+**Seed 123**
+
+| BF16 · Click to open original PNG | INT8 · Click to open original PNG |
+|---|---|
+| [![Image editing: sweater recoloring · BF16 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/edit-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/bf16/edit-s123.png) | [![Image editing: sweater recoloring · INT8 · seed 123](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/edit-s123.png)](https://huggingface.co/ixim/Image21-INT8/resolve/main/evaluation/int8/edit-s123.png) |
+
+SHA256 prefixes: BF16 `7d00c5bb7a2b` / INT8 `d8023e34ba10`.
+
+### Machine-readable evidence
+
+The measured data and all images above are embedded in this card. Original files remain available for independent checking:
+
+- [Core CSV](evaluation/comparison.csv) · [Portrait CSV](evaluation/young_woman/comparison.csv)
+- [Core BF16 JSONL](evaluation/bf16/records.jsonl) · [Core INT8 JSONL](evaluation/int8/records.jsonl)
+- [Portrait BF16 JSONL](evaluation/young_woman/bf16/records.jsonl) · [Portrait INT8 JSONL](evaluation/young_woman/int8/records.jsonl)
+- [Core environment](evaluation/bf16/environment.json) · [INT8 environment](evaluation/int8/environment.json)
+- [File checksums](MANIFEST.json) · [Quantization provenance](conversion.json)
+
+
+## License and modifications
+
+The weights remain subject to the [Qwen Research License Agreement](LICENSE):
+non-commercial research and evaluation only; commercial use requires a separate
+license from the original licensor. Redistribution must preserve that agreement,
+attribution and modification notices. See [Notice](Notice) and [CHANGES.md](CHANGES.md).
+
+Source: [Hugging Face](https://huggingface.co/Qwen/Qwen-Image-2.1) |
+[ModelScope](https://modelscope.cn/models/Qwen/Qwen-Image-2.1).

REPRODUCE.md

@@ -0,0 +1,113 @@
+# Image21-INT8 发布工作区
+
+目标：为 Qwen-Image-2.1 制作可重新加载的 Diffusers INT8 衍生模型，比较 BF16 与
+INT8 实际效果，并准备 Hugging Face（ixim）英文、ModelScope（iximbox）中文发布包。
+
+本地下载、量化和原有 28 张正式评测输出已完成；另增加四张中国年轻成年女性肖像
+补测输出。两平台目标名称统一为 Image21-INT8。Hugging Face 登录账号已核验为 ixim。
+最新上传状态及 ModelScope 本地登录方法见 [PUBLISHING.md](PUBLISHING.md)。
+
+工作区内的 `cards/` 是模板，
+`release/` 中通过验证的目录才是上传来源；没有实测记录时不能宣称发布完成。
+
+主名称使用 **Image21-INT8**，模型说明中描述其 Qwen-Image-2.1 来源。
+原模型采用 Qwen Research License，仅限非商业研究与评测。
+
+## 环境
+
+Windows PowerShell，新机器建议创建不继承全局软件包的虚拟环境：
+
+```powershell
+python -m venv .venv
+.venv/Scripts/python.exe -m pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
+.venv/Scripts/python.exe -m pip install -r requirements.txt
+```
+
+本次工作区为复用已有 CUDA PyTorch，使用了 `--system-site-packages`；更新仅安装
+在 `.venv` 内。脚本将 Numba 和 Hugging Face 缓存定向到项目 `.cache`。
+RTX 5090 可通过模型 CPU 卸载运行；精确显存需求以实测报告为准。
+
+## 下载、量化、测评
+
+在工作区根目录执行：
+
+```powershell
+.venv/Scripts/python.exe -m scripts.download --source modelscope
+.venv/Scripts/python.exe -m scripts.integrity
+.venv/Scripts/python.exe -m scripts.quantize
+.venv/Scripts/python.exe -m scripts.benchmark --model models/bf16 --output artifacts/eval/bf16
+.venv/Scripts/python.exe -m scripts.benchmark --model models/int8 --output artifacts/eval/int8
+```
+
+下载固定版本配置，ModelScope 权重按固定 HF 版本 SHA256 核验。下载不使用第三方镜像。
+量化保留敏感层和 VAE 浮点精度，具体模块记录在 `*-quantization.json`。
+
+日常本地生成：
+
+```powershell
+.venv/Scripts/python.exe -m scripts.infer --model models/int8 --prompt '一只坐在窗边的橘猫，暖色摄影' --output outputs/cat.png
+```
+
+编辑时增加 `--input input.png`；多图可在该参数后继续添加文件名。
+默认 1024×1024、40 步、CPU 卸载，PNG 保留 RGBA 通道。
+
+输出目录非空时量化拒绝覆盖；测评结果目录已有记录时拒绝覆盖。
+编辑用例输入取自 BF16 的 portrait-s42，因此必须先完成 BF16 测评。
+默认 1024×1024、40 步、种子 42/123、相同卸载方式，另做一次不计入的完整预热。
+
+肖像补测单独运行、单独统计，同样先完成 BF16 再运行 INT8：
+
+```powershell
+.venv/Scripts/python.exe -m scripts.benchmark --model models/bf16 --output artifacts/eval/young_woman/bf16 --cases benchmarks/young_woman.json
+.venv/Scripts/python.exe -m scripts.benchmark --model models/int8 --output artifacts/eval/young_woman/int8 --cases benchmarks/young_woman.json
+.venv/Scripts/python.exe -m scripts.report --baseline artifacts/eval/young_woman/bf16 --quantized artifacts/eval/young_woman/int8 --output artifacts/eval/young_woman
+```
+
+快速端到端检查（不能作为质量评测）：
+
+```powershell
+.venv/Scripts/python.exe -m scripts.benchmark --model models/bf16 --output artifacts/smoke/bf16 --case portrait --seeds 42 --width 512 --height 512 --steps 2 --no-warmup
+```
+
+## 校验
+
+```powershell
+.venv/Scripts/python.exe -m unittest discover -s tests -v
+$env:NUMBA_CACHE_DIR = "$PWD/.cache/numba"
+.venv/Scripts/python.exe -m tests.gpu_roundtrip
+```
+
+保留 `artifacts/upstream` 官方清单及许可证、下载校验记录、评测原始记录和图片。
+不要将 `.cache`、`.venv`、账户 Token 或 BF16 源模型目录上传到 INT8 仓库。
+
+## 生成发布文件与上传
+
+```powershell
+.venv/Scripts/python.exe -m scripts.release annotate
+.venv/Scripts/python.exe -m scripts.report
+.venv/Scripts/python.exe -m scripts.release stage --platform huggingface --output release/huggingface
+.venv/Scripts/python.exe -m scripts.release stage --platform modelscope --output release/modelscope
+.venv/Scripts/python.exe -m scripts.upload huggingface --dry-run
+.venv/Scripts/python.exe -m scripts.upload modelscope --dry-run
+```
+
+`annotate` 在量化文件中写入许可证要求的修改声明；建议在 INT8 实测前执行。
+`stage` 要求完整正式评测，生成平台 README、许可证、归属、脚本、评测图片及逐文件清单。
+上传前会再次核验全部文件，不上传清单外内容。最终效果可打开
+`artifacts/eval/comparison.html` 逐图检查。
+
+Hugging Face 在本机交互登录（Token 不要粘贴到聊天）：
+
+```powershell
+$env:HF_HOME = "$PWD/.cache/huggingface"
+.venv/Scripts/hf.exe auth login
+.venv/Scripts/python.exe -m scripts.upload huggingface
+```
+
+ModelScope 使用已登录账号。需要登录时，以下交互式 Python 输入不回显，
+也不会把 Token 放入命令历史：
+
+```powershell
+.venv/Scripts/python.exe -c "from getpass import getpass; from modelscope.hub.api import HubApi; HubApi().login(getpass('ModelScope token: '))"
+.venv/Scripts/python.exe -m scripts.upload modelscope
+```

benchmarks/cases.json

@@ -0,0 +1,9 @@
+[
+  {"id":"portrait","category":"portrait","prompt":"A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition."},
+  {"id":"english_text","category":"typography_en","prompt":"A clean editorial poster with a deep blue background. Large exact headline at the top: \"CREATE WITH LIGHT\". Smaller exact subtitle: \"September 2026\". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text."},
+  {"id":"chinese_text","category":"typography_zh","prompt":"设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。"},
+  {"id":"composition","category":"spatial_counting","prompt":"A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text."},
+  {"id":"texture","category":"texture","prompt":"Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures."},
+  {"id":"rgba","category":"transparency","prompt":"This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent."},
+  {"id":"edit","category":"image_editing","prompt":"Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.","input":"artifacts/eval/bf16/portrait-s42.png"}
+]

benchmarks/young_woman.json

@@ -0,0 +1,7 @@
+[
+  {
+    "id": "young_chinese_woman",
+    "category": "adult_portrait",
+    "prompt": "一张写实摄影风格的中国年轻成年女性肖像，人物明确为25岁成年人。黑色及肩直发，穿着米白色圆领针织衫，胸部以上构图，正面看向镜头，神情自然放松，轻微微笑。窗边柔和的自然日光，浅灰色简洁背景，真实自然的皮肤纹理和少量雀斑，淡妆，不过度磨皮。85mm人像摄影，浅景深，色彩自然，画面中没有文字。"
+  }
+]

cards/huggingface.md

@@ -0,0 +1,107 @@
+---
+license: other
+license_name: qwen-research
+license_link: LICENSE
+base_model: Qwen/Qwen-Image-2.1
+base_model_relation: quantized
+library_name: diffusers
+pipeline_tag: text-to-image
+language:
+  - en
+  - zh
+tags:
+  - diffusers
+  - bitsandbytes
+  - int8
+  - image-generation
+  - image-editing
+  - rgba
+---
+
+# Image21-INT8
+
+**Built with Qwen.** An independent bitsandbytes INT8 derivative of
+[Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1), prepared by **ixim**.
+This is a community conversion, not an official Qwen release.
+
+## Quantization
+
+Base revision: `b3179ad355be050328e483a9dfdd9e60cd62adfa`.
+Eligible linear layers in the generation transformer and the Qwen3-VL text encoder
+use **bitsandbytes LLM.int8**, with outlier threshold **6.0**. This is mixed-precision
+inference: outlier handling uses floating point; the VAE, normalization, embeddings,
+vision encoder, output head, and excluded conditioning/projection layers remain
+floating point. See the two component quantization JSON files for exact coverage.
+No fine-tuning or calibration dataset was used. This is not a pure W8A8 pipeline,
+FP8 conversion, or GGUF checkpoint. Lower storage does not imply faster inference.
+
+## Installation and loading
+
+Use Python 3.13 and a CUDA-capable PyTorch installation. The tested runtime uses
+PyTorch 2.10.0+cu128 on an RTX 5090. Install the accompanying `requirements.txt`;
+it pins Diffusers to the official commit used for this conversion.
+
+```bash
+pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
+pip install -r requirements.txt
+```
+
+```python
+import torch
+from diffusers import QwenImage21Pipeline
+
+pipe = QwenImage21Pipeline.from_pretrained(
+    "ixim/Image21-INT8", dtype=torch.bfloat16,
+)
+pipe.enable_model_cpu_offload()
+image = pipe(
+    prompt='A neon sign reading "CREATE WITH LIGHT", rainy night',
+    width=1024, height=1024, num_inference_steps=40,
+    true_cfg_scale=1.0, use_kv_cache=True,
+    generator=torch.Generator("cpu").manual_seed(42),
+).images[0]
+image.save("output.png")
+```
+
+The saved quantization configuration is detected automatically. Do not re-quantize
+these weights at load time. For editing, pass `image=PIL.Image.open("input.png")`.
+For transparency, use the upstream RGBA prompt format and save PNG files.
+
+## Local evaluation
+
+{{EVALUATION_EN}}
+
+**Observed trade-off:** weights are smaller and average latency is lower in this
+run, but peak allocated GPU memory is **higher** with the tested INT8/offload stack.
+Do not use this checkpoint as evidence of lower-VRAM operation.
+
+**Editing limitation:** both BF16 and INT8 sweater-edit examples show substantial
+oversharpening and contrast changes. Faithful preservation of the original image
+is not established. Both precisions also add unwanted footer text in the inspected
+Chinese poster. See [qualitative observations](evaluation/qualitative.md).
+
+The comparison uses identical prompts, CPU random generators, seeds, resolutions,
+40 denoising steps, CFG=1, KV cache enabled, and model CPU offload. Warmup is excluded.
+Peak allocated/reserved memory are PyTorch CUDA statistics, not total board memory.
+Timing includes prompt encoding, denoising, decoding and CPU transfers; it excludes
+model loading and image-file writing. Paired pixel differences measure drift, not
+semantic quality. This small suite is not a standardized quality ranking.
+BF16 was run first, followed by INT8 on the same Windows desktop GPU; repeated
+timing trials and confidence intervals were not collected.
+
+See [evaluation/report.md](evaluation/report.md), raw JSONL records, environment
+records and paired PNG outputs. The edit pair uses the exact same source image.
+Higher resolutions, extensive multi-reference editing, other GPUs, other attention
+backends, and ComfyUI compatibility are not established by this evaluation.
+
+{{EVALUATION_DETAILS_EN}}
+
+## License and modifications
+
+The weights remain subject to the [Qwen Research License Agreement](LICENSE):
+non-commercial research and evaluation only; commercial use requires a separate
+license from the original licensor. Redistribution must preserve that agreement,
+attribution and modification notices. See [Notice](Notice) and [CHANGES.md](CHANGES.md).
+
+Source: [Hugging Face](https://huggingface.co/Qwen/Qwen-Image-2.1) |
+[ModelScope](https://modelscope.cn/models/Qwen/Qwen-Image-2.1).

cards/modelscope.md

@@ -0,0 +1,100 @@
+---
+license: other
+license_name: qwen-research
+license_link: LICENSE
+base_model: Qwen/Qwen-Image-2.1
+base_model_relation: quantized
+library_name: diffusers
+pipeline_tag: text-to-image
+language:
+  - zh
+  - en
+tags:
+  - diffusers
+  - bitsandbytes
+  - int8
+  - image-generation
+  - image-editing
+  - rgba
+---
+
+# Image21-INT8
+
+**Built with Qwen。** 由 **iximbox** 制作的
+[Qwen-Image-2.1](https://modelscope.cn/models/Qwen/Qwen-Image-2.1)
+社区 INT8 量化衍生版本，面向 Diffusers / Python 推理，不属于 Qwen 官方发布。
+
+## 量化说明
+
+基模型固定版本：`b3179ad355be050328e483a9dfdd9e60cd62adfa`（Hugging Face）。
+两平台官方权重经过 SHA256 一致性核验。
+
+生成 Transformer 与 Qwen3-VL 文本编码器中适用的线性层采用
+**bitsandbytes LLM.int8**，离群值阈值为 **6.0**。这属于混合精度推理：
+离群值计算使用浮点精度，VAE、归一化、嵌入、视觉编码器、输出头及排除的
+条件投影层保留浮点精度。具体覆盖范围见两个组件的量化 JSON 报告。
+未进行微调，无需校准数据集。此模型不是全链路 W8A8、FP8 或 GGUF 模型。
+体积减小不等于推理加速，速度以实际测评为准。
+
+## 安装与使用
+
+测试环境为 Python 3.13、PyTorch 2.10.0+cu128 和 RTX 5090。
+`requirements.txt` 固定了本次验证使用的官方 Diffusers 提交及主要依赖。
+
+```bash
+pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu128
+pip install -r requirements.txt
+```
+
+```python
+import torch
+from modelscope import snapshot_download
+from diffusers import QwenImage21Pipeline
+
+model_dir = snapshot_download("iximbox/Image21-INT8")
+pipe = QwenImage21Pipeline.from_pretrained(model_dir, dtype=torch.bfloat16)
+pipe.enable_model_cpu_offload()
+image = pipe(
+    prompt='一张咖啡馆海报，清晰写着“慢下来，喝杯咖啡”，暖色摄影',
+    width=1024, height=1024, num_inference_steps=40,
+    true_cfg_scale=1.0, use_kv_cache=True,
+    generator=torch.Generator("cpu").manual_seed(42),
+).images[0]
+image.save("output.png")
+```
+
+加载时会自动读取已保存的量化配置，无需再次指定量化参数。
+图像编辑可增加 `image=PIL.Image.open("input.png")`。
+透明图生成使用官方 RGBA 提示词格式，并保存为 PNG 以保留透明通道。
+
+## 实测对比
+
+{{EVALUATION_ZH}}
+
+**实际取舍：** 本次测试中权重更小、平均耗时更低，但 INT8 搭配当前卸载方案的
+已分配显存峰值**更高**。不能据此宣称本版本适用于更低显存设备。
+
+**编辑限制：** BF16 与 INT8 的毛衣改色样例都出现明显过度锐化及对比度变化，
+未证明能够忠实保持原图质感。所检查的中文海报在两种精度中也均出现了多余页脚文字。
+详见 [定性检查](evaluation/qualitative.md)。
+
+两组使用相同提示词、CPU 随机数生成器、种子、分辨率、40 步、CFG=1、KV 缓存
+及按模型 CPU 卸载。预热不计入结果。显存数值为 PyTorch CUDA 分配/保留峰值，
+不代表显卡总占用。耗时包含编码、去噪、解码及 CPU 传输，不包含加载和图片写盘。
+像素差异用于观察量化偏移，不能等同于语义质量评分；小样本不能代表标准榜单成绩。
+测试在同一 Windows 桌面显卡上先运行 BF16，再运行 INT8，未做多轮独立计时或置信区间估计。
+
+详见 [评测报告](evaluation/report.md)、原始 JSONL、环境记录与成对 PNG。
+编辑测试使用同一份输入图像。更高分辨率、大量多图参考、其他显卡、其他注意力
+后端及 ComfyUI 兼容性未由本次评测证明。
+
+{{EVALUATION_DETAILS_ZH}}
+
+## 许可证与修改声明
+
+权重沿用 [Qwen Research License Agreement](LICENSE)，仅用于非商业研究与评测；
+商业使用需向原许可方另行申请。再分发须保留原协议、归属及修改说明。
+详见 [Notice](Notice) 和 [CHANGES.md](CHANGES.md)。
+
+原始模型：[ModelScope](https://modelscope.cn/models/Qwen/Qwen-Image-2.1) |
+[Hugging Face](https://huggingface.co/Qwen/Qwen-Image-2.1)。

conversion.json

@@ -0,0 +1,39 @@
+{
+  "base_model": "Qwen/Qwen-Image-2.1",
+  "base_revision": "b3179ad355be050328e483a9dfdd9e60cd62adfa",
+  "diffusers_commit": "80c7ed262aeffbeb43ef13ae04baeb9b84515a69",
+  "method": "bitsandbytes LLM.int8",
+  "threshold": 6.0,
+  "weight_files": [
+    {
+      "path": "text_encoder/model-00001-of-00003.safetensors",
+      "size": 4999822155,
+      "sha256": "9b8ae3bd2a1e4884abc48c659212427bbbae06fb1bc17e7173583704fc0f6bdd"
+    },
+    {
+      "path": "text_encoder/model-00002-of-00003.safetensors",
+      "size": 4996432636,
+      "sha256": "d28d1a8b3994a43582acc5e2da15ff38b632fae1466f41bc7710d27ec170aa9e"
+    },
+    {
+      "path": "text_encoder/model-00003-of-00003.safetensors",
+      "size": 27729689,
+      "sha256": "5896ffa9e2b2a52f9d496755fde050306f21c449bd66c2526b40979033c6e244"
+    },
+    {
+      "path": "transformer/diffusion_pytorch_model-00001-of-00002.safetensors",
+      "size": 4989447305,
+      "sha256": "0d77daaa3129860f026567e0178fec4ed776f3cf2463da76a0823b9c381c3468"
+    },
+    {
+      "path": "transformer/diffusion_pytorch_model-00002-of-00002.safetensors",
+      "size": 2267331623,
+      "sha256": "5e1b8ee3917833687adb662830438dc612ace2cf4bff7b8898515184a888fde1"
+    },
+    {
+      "path": "vae/diffusion_pytorch_model.safetensors",
+      "size": 1350989512,
+      "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+    }
+  ]
+}

evaluation/bf16/environment.json

@@ -0,0 +1,155 @@
+{
+  "python": "3.13.5",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "gpu": "NVIDIA GeForce RTX 5090",
+  "cuda": "12.8",
+  "packages": {
+    "torch": "2.10.0+cu128",
+    "torchvision": "0.25.0+cu128",
+    "diffusers": "0.41.0.dev0",
+    "transformers": "5.17.0",
+    "accelerate": "1.15.0",
+    "bitsandbytes": "0.50.2",
+    "safetensors": "0.8.0",
+    "huggingface-hub": "1.32.0",
+    "tokenizers": "0.23.2",
+    "numpy": "2.2.6",
+    "pillow": "12.2.0",
+    "psutil": "7.0.0"
+  },
+  "model": "D:\\Git\\Qwen-Image-2.1\\models\\bf16",
+  "load_seconds": 10.8723844000001,
+  "load_peak_allocated_bytes": 0,
+  "offload": "model",
+  "warmup": true,
+  "generator_device": "cpu",
+  "cases_sha256": "2663ee3d489e1eb136e02b083963b8622146fc65ac301d6acc5a6aab21e63356",
+  "model_identity": {
+    "kind": "bf16",
+    "base_revision": "b3179ad355be050328e483a9dfdd9e60cd62adfa",
+    "files": [
+      {
+        "path": "model_index.json",
+        "sha256": "cf1ecd104ea090855d60d8cec0895c1e9d6ee41b2f87e231b678beecd1cf7809",
+        "size": 447
+      },
+      {
+        "path": "processor/added_tokens.json",
+        "sha256": "c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680",
+        "size": 707
+      },
+      {
+        "path": "processor/chat_template.jinja",
+        "sha256": "3636d0f0bd6bef02654cdffdc447b79cb2cef8ab02cc75267345946291a489e4",
+        "size": 5292
+      },
+      {
+        "path": "processor/merges.txt",
+        "sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5",
+        "size": 1671853
+      },
+      {
+        "path": "processor/preprocessor_config.json",
+        "sha256": "93585062a80db5e8ca038efc7726a3e6411d9db948472d81d63c6303993be8c5",
+        "size": 782
+      },
+      {
+        "path": "processor/special_tokens_map.json",
+        "sha256": "76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd",
+        "size": 613
+      },
+      {
+        "path": "processor/tokenizer.json",
+        "sha256": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4",
+        "size": 11422654
+      },
+      {
+        "path": "processor/tokenizer_config.json",
+        "sha256": "81ec7bb9530159b326c0bef1d0b6c33d392090524014ea3f0123a3c1eb9c2af5",
+        "size": 5445
+      },
+      {
+        "path": "processor/video_preprocessor_config.json",
+        "sha256": "59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1",
+        "size": 817
+      },
+      {
+        "path": "processor/vocab.json",
+        "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910",
+        "size": 2776833
+      },
+      {
+        "path": "scheduler/scheduler_config.json",
+        "sha256": "5895f3a167c14a967fe9ac70c64924ae5acc79799e0679fd12907e594a713cd1",
+        "size": 485
+      },
+      {
+        "path": "text_encoder/config.json",
+        "sha256": "6001331949f7f86ab6d791a80b12246d4f8da5159fa08c3e62368a39f36d63a9",
+        "size": 1517
+      },
+      {
+        "path": "text_encoder/generation_config.json",
+        "sha256": "4d9818c3d27895c0058828a5f68bc7a4de80c3ae1bcdac90936ce24178063f59",
+        "size": 213
+      },
+      {
+        "path": "text_encoder/model-00001-of-00004.safetensors",
+        "size": 4998056552,
+        "sha256": "dde00291b5f7fb92013895310a3da0ddba78674df9f10d505d375243dc01fc6f"
+      },
+      {
+        "path": "text_encoder/model-00002-of-00004.safetensors",
+        "size": 4915962464,
+        "sha256": "9047faccc0a6d98496a52d55f27be1c94a9c259d1e283fbea0128d054a948d42"
+      },
+      {
+        "path": "text_encoder/model-00003-of-00004.safetensors",
+        "size": 4915962496,
+        "sha256": "8c54187654c0176b73ae73785bf791dc9a14c9df7fb4310083a09d42048cb57e"
+      },
+      {
+        "path": "text_encoder/model-00004-of-00004.safetensors",
+        "size": 2704357976,
+        "sha256": "5311532aaaeae3259eb6a7b2c600636be1159adf7ded35f53579f7d0e7d43cdd"
+      },
+      {
+        "path": "text_encoder/model.safetensors.index.json",
+        "sha256": "859146bff8db1d617fca8b457a5223edff963bb70076b3645f2189fe0540a147",
+        "size": 67795
+      },
+      {
+        "path": "transformer/config.json",
+        "sha256": "56ae3281c4e6c2d1aa3658252d187488071815fd79bef15808bb0205fc1a2241",
+        "size": 370
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model-00001-of-00002.safetensors",
+        "size": 9968332504,
+        "sha256": "9e6bc2d641e67bf277895ea8777141044a38f3edb7101bc469b2961dd7c36b4b"
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model-00002-of-00002.safetensors",
+        "size": 4261951904,
+        "sha256": "3aaf234dcbe128530479735854a346b5e3e66283b7c11db56f836bbd1c13ebaa"
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model.safetensors.index.json",
+        "sha256": "17987f6623b1c814d0ef55a137d99142b7b3b040eb1bf241b5575dd35af803a2",
+        "size": 30283
+      },
+      {
+        "path": "vae/config.json",
+        "sha256": "9785d527b278cb8b210e9f48a8028d92a4190968ce7e24c8fc85d9d1b82f6ba2",
+        "size": 2079
+      },
+      {
+        "path": "vae/diffusion_pytorch_model.safetensors",
+        "size": 1350989512,
+        "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+      }
+    ],
+    "fingerprint": "74c96a68092fdb1f3eed2063bcfba8fd99380ada6234e405e646681f35fb6984"
+  },
+  "model_identity_origin": "Bound during run to official weight and config manifests independently verified before model loading; file hashes will be rechecked after evaluation."
+}

evaluation/bf16/records.jsonl

@@ -0,0 +1,14 @@
+{"case_id": "portrait", "category": "portrait", "prompt": "A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 29.766559700000016, "peak_allocated_bytes": 17611056128, "peak_reserved_bytes": 17800626176, "process_rss_after_bytes": 42226987008, "image": "portrait-s42.png", "image_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 239, "alpha_max": 255, "alpha_nonopaque_fraction": 0.015043258666992188}
+{"case_id": "portrait", "category": "portrait", "prompt": "A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 26.54884089999996, "peak_allocated_bytes": 17611056128, "peak_reserved_bytes": 17800626176, "process_rss_after_bytes": 45726285824, "image": "portrait-s123.png", "image_sha256": "cff376e50c41c4aee2b9e4aabc502f4791774f6ad285d8a7673a61ebb1fa650d", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 252, "alpha_max": 255, "alpha_nonopaque_fraction": 0.07719135284423828}
+{"case_id": "english_text", "category": "typography_en", "prompt": "A clean editorial poster with a deep blue background. Large exact headline at the top: \"CREATE WITH LIGHT\". Smaller exact subtitle: \"September 2026\". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.952576900000167, "peak_allocated_bytes": 17619124224, "peak_reserved_bytes": 17806917632, "process_rss_after_bytes": 46901248000, "image": "english_text-s42.png", "image_sha256": "ba31b0c8a312e96408096566013c667b3a32bd7720749ca47120960b17b41ca6", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 250, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0008878707885742188}
+{"case_id": "english_text", "category": "typography_en", "prompt": "A clean editorial poster with a deep blue background. Large exact headline at the top: \"CREATE WITH LIGHT\". Smaller exact subtitle: \"September 2026\". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.878974200000357, "peak_allocated_bytes": 17619124224, "peak_reserved_bytes": 17806917632, "process_rss_after_bytes": 48110997504, "image": "english_text-s123.png", "image_sha256": "cf5e76dee9f0e403ebdf24a64e1fcd604ee2b6c8e645a032fba2807acc34adf2", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 221, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0014019012451171875}
+{"case_id": "chinese_text", "category": "typography_zh", "prompt": "设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.469736900000044, "peak_allocated_bytes": 17623597568, "peak_reserved_bytes": 17813209088, "process_rss_after_bytes": 48624263168, "image": "chinese_text-s42.png", "image_sha256": "7082cb71d99705bc42e27e2c206b13e3beb9e62cfd6304cfabe203b82641692c", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 245, "alpha_max": 255, "alpha_nonopaque_fraction": 0.017661094665527344}
+{"case_id": "chinese_text", "category": "typography_zh", "prompt": "设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.84774830000015, "peak_allocated_bytes": 17623597568, "peak_reserved_bytes": 17813209088, "process_rss_after_bytes": 49212678144, "image": "chinese_text-s123.png", "image_sha256": "eae74256d7ed1b3519d39c7d58961cf7a937224d19a3c3239ca7eba4178edf49", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 248, "alpha_max": 255, "alpha_nonopaque_fraction": 0.013823509216308594}
+{"case_id": "composition", "category": "spatial_counting", "prompt": "A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.38054239999974, "peak_allocated_bytes": 17621962752, "peak_reserved_bytes": 17813209088, "process_rss_after_bytes": 49377714176, "image": "composition-s42.png", "image_sha256": "91e967685fbbd79ba13daece22a0f72c863cb518db88d5d2a6f9aecfc5f292b1", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 252, "alpha_max": 255, "alpha_nonopaque_fraction": 0.005370140075683594}
+{"case_id": "composition", "category": "spatial_counting", "prompt": "A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.351097499999923, "peak_allocated_bytes": 17621962752, "peak_reserved_bytes": 17813209088, "process_rss_after_bytes": 49788104704, "image": "composition-s123.png", "image_sha256": "bc11612f260acd8af0d834f3b28c752cefd0437d24c9a01dae0d46647bdf29ae", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 252, "alpha_max": 255, "alpha_nonopaque_fraction": 0.009993553161621094}
+{"case_id": "texture", "category": "texture", "prompt": "Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.850519100000383, "peak_allocated_bytes": 17607326208, "peak_reserved_bytes": 17798529024, "process_rss_after_bytes": 49929191424, "image": "texture-s42.png", "image_sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 251, "alpha_max": 255, "alpha_nonopaque_fraction": 0.03330421447753906}
+{"case_id": "texture", "category": "texture", "prompt": "Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.7564997999998, "peak_allocated_bytes": 17607326208, "peak_reserved_bytes": 17798529024, "process_rss_after_bytes": 50300633088, "image": "texture-s123.png", "image_sha256": "18c4403f388d0e03447473c03497141169f333fddeaeb40227163dc3b9d68d3e", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 251, "alpha_max": 255, "alpha_nonopaque_fraction": 0.016788482666015625}
+{"case_id": "rgba", "category": "transparency", "prompt": "This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 26.040632799999912, "peak_allocated_bytes": 17604822528, "peak_reserved_bytes": 17798529024, "process_rss_after_bytes": 51127226368, "image": "rgba-s42.png", "image_sha256": "6f5dfea6ba02e41cbf20617b6e863e1adfa139d79fad2ca5c64dfc815f591d82", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 0, "alpha_max": 255, "alpha_nonopaque_fraction": 0.766596794128418}
+{"case_id": "rgba", "category": "transparency", "prompt": "This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 25.074647800000093, "peak_allocated_bytes": 17604822528, "peak_reserved_bytes": 17798529024, "process_rss_after_bytes": 50980712448, "image": "rgba-s123.png", "image_sha256": "6f1abbb1a64ac553fc576a0250d370ab5bb7e6d5e6372f42f67045a5bb0951a6", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 0, "alpha_max": 255, "alpha_nonopaque_fraction": 0.6571817398071289}
+{"case_id": "edit", "category": "image_editing", "prompt": "Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa", "seconds": 31.15521630000012, "peak_allocated_bytes": 20488328192, "peak_reserved_bytes": 21506293760, "process_rss_after_bytes": 51628781568, "image": "edit-s42.png", "image_sha256": "8713f6ef2ec0191d88225f2515671d6f51be43aa293bb1ca6e25b07a1b67deb9", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 228, "alpha_max": 255, "alpha_nonopaque_fraction": 0.11233901977539062}
+{"case_id": "edit", "category": "image_editing", "prompt": "Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa", "seconds": 31.030128800000057, "peak_allocated_bytes": 20488328192, "peak_reserved_bytes": 21506293760, "process_rss_after_bytes": 52064702464, "image": "edit-s123.png", "image_sha256": "7d00c5bb7a2b31f05ea3fcf3f4a79dd93335770fe1b2d0ad772d5df954c9f91c", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 253, "alpha_max": 255, "alpha_nonopaque_fraction": 0.010979652404785156}

evaluation/comparison.csv

@@ -0,0 +1,15 @@
+case_id,seed,bf16_seconds,int8_seconds,speed_ratio_bf16_over_int8,bf16_peak_allocated_gib,int8_peak_allocated_gib,rgb_mae_white,rgb_psnr_white_db,alpha_mae
+portrait,42,29.766559700000016,19.65371929999992,1.5145509735656058,16.40157413482666,22.8613600730896,0.013405879028141499,31.702880072340637,4.5881097321398556e-05
+portrait,123,26.54884089999996,19.614630900000066,1.3535223290895506,16.40157413482666,22.86135959625244,0.0162954181432724,30.546014578484698,0.0001803267077775672
+english_text,42,25.952576900000167,19.416275399999904,1.3366403373120828,16.409088134765625,22.861863613128662,0.026258574798703194,19.447587010373717,8.03704188001575e-06
+english_text,123,25.878974200000357,19.404967399999805,1.3336262652005597,16.409088134765625,22.86188554763794,0.011798395775258541,25.825570953357694,1.0239845323667396e-05
+chinese_text,42,25.469736900000044,19.544581299999663,1.303161040344235,16.413254261016846,22.862136363983154,0.05972025915980339,15.60138057172976,0.0001082326634787023
+chinese_text,123,25.84774830000015,19.86057179999989,1.3014604292511,16.413254261016846,22.86215353012085,0.024602750316262245,20.44133777040833,6.83055113768205e-05
+composition,42,25.38054239999974,19.130140400000073,1.326730586880567,16.411731719970703,22.862086296081543,0.02268194407224655,23.666249907034246,2.1964429834042676e-05
+composition,123,25.351097499999923,19.303240899999764,1.3133078342300661,16.411731719970703,22.86205244064331,0.01959226094186306,22.865933886645216,3.4590328141348436e-05
+texture,42,25.850519100000383,19.223099399999683,1.3447633267713743,16.39810037612915,22.86104154586792,0.06978315860033035,18.01562989780135,0.00019098912889603525
+texture,123,25.7564997999998,19.567528400000356,1.3162878455308302,16.39810037612915,22.861046314239502,0.01683949865400791,27.92905075564026,5.856681673321873e-05
+rgba,42,26.040632799999912,19.388802900000428,1.3430758430165555,16.395768642425537,22.860955238342285,0.018291756510734558,20.893335328149618,0.01108772773295641
+rgba,123,25.074647800000093,19.288307799999984,1.2999921019510128,16.395768642425537,22.86095952987671,0.04646634683012962,17.051910336330792,0.029543064534664154
+edit,42,31.15521630000012,25.399254500000097,1.2266193206576201,19.081242561340332,23.079509735107422,0.009711389429867268,36.48036013354599,0.00011113862274214625
+edit,123,31.030128800000057,25.24017350000031,1.2293944334415798,19.081242561340332,23.080156326293945,0.003908749669790268,43.267885143031855,2.616058372950647e-05

evaluation/comparison.html

@@ -0,0 +1,18 @@
+<!doctype html><meta charset="utf-8"><title>BF16 / INT8 comparison</title>
+<style>body{font:16px system-ui;margin:32px;max-width:1500px}section{margin:40px 0}.pair{display:grid;grid-template-columns:1fr 1fr;gap:20px}img{width:100%;background:repeating-conic-gradient(#ddd 0 25%,#fff 0 50%) 0/24px 24px}pre{white-space:pre-wrap}</style>
+<h1>BF16 / INT8 comparison</h1><p>Left: BF16. Right: saved INT8. Identical settings and seeds. Alpha displayed over checkerboard.</p>
+<p><a href="young_woman/comparison.html">Additional adult Chinese woman portrait pairs</a></p>
+<section><h2>portrait — seed 42</h2><pre>A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.</pre><div class="pair"><img src="bf16/portrait-s42.png" alt="BF16"><img src="int8/portrait-s42.png" alt="INT8"></div></section>
+<section><h2>portrait — seed 123</h2><pre>A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.</pre><div class="pair"><img src="bf16/portrait-s123.png" alt="BF16"><img src="int8/portrait-s123.png" alt="INT8"></div></section>
+<section><h2>english_text — seed 42</h2><pre>A clean editorial poster with a deep blue background. Large exact headline at the top: &quot;CREATE WITH LIGHT&quot;. Smaller exact subtitle: &quot;September 2026&quot;. A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.</pre><div class="pair"><img src="bf16/english_text-s42.png" alt="BF16"><img src="int8/english_text-s42.png" alt="INT8"></div></section>
+<section><h2>english_text — seed 123</h2><pre>A clean editorial poster with a deep blue background. Large exact headline at the top: &quot;CREATE WITH LIGHT&quot;. Smaller exact subtitle: &quot;September 2026&quot;. A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.</pre><div class="pair"><img src="bf16/english_text-s123.png" alt="BF16"><img src="int8/english_text-s123.png" alt="INT8"></div></section>
+<section><h2>chinese_text — seed 42</h2><pre>设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。</pre><div class="pair"><img src="bf16/chinese_text-s42.png" alt="BF16"><img src="int8/chinese_text-s42.png" alt="INT8"></div></section>
+<section><h2>chinese_text — seed 123</h2><pre>设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。</pre><div class="pair"><img src="bf16/chinese_text-s123.png" alt="BF16"><img src="int8/chinese_text-s123.png" alt="INT8"></div></section>
+<section><h2>composition — seed 42</h2><pre>A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.</pre><div class="pair"><img src="bf16/composition-s42.png" alt="BF16"><img src="int8/composition-s42.png" alt="INT8"></div></section>
+<section><h2>composition — seed 123</h2><pre>A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.</pre><div class="pair"><img src="bf16/composition-s123.png" alt="BF16"><img src="int8/composition-s123.png" alt="INT8"></div></section>
+<section><h2>texture — seed 42</h2><pre>Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.</pre><div class="pair"><img src="bf16/texture-s42.png" alt="BF16"><img src="int8/texture-s42.png" alt="INT8"></div></section>
+<section><h2>texture — seed 123</h2><pre>Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.</pre><div class="pair"><img src="bf16/texture-s123.png" alt="BF16"><img src="int8/texture-s123.png" alt="INT8"></div></section>
+<section><h2>rgba — seed 42</h2><pre>This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.</pre><div class="pair"><img src="bf16/rgba-s42.png" alt="BF16"><img src="int8/rgba-s42.png" alt="INT8"></div></section>
+<section><h2>rgba — seed 123</h2><pre>This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.</pre><div class="pair"><img src="bf16/rgba-s123.png" alt="BF16"><img src="int8/rgba-s123.png" alt="INT8"></div></section>
+<section><h2>edit — seed 42</h2><pre>Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.</pre><div class="pair"><img src="bf16/edit-s42.png" alt="BF16"><img src="int8/edit-s42.png" alt="INT8"></div></section>
+<section><h2>edit — seed 123</h2><pre>Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.</pre><div class="pair"><img src="bf16/edit-s123.png" alt="BF16"><img src="int8/edit-s123.png" alt="INT8"></div></section>

evaluation/hardware.json

@@ -0,0 +1,10 @@
+{
+  "gpu_driver_memory": "NVIDIA GeForce RTX 5090, 616.92, 32607",
+  "system_ram_bytes": 274298757120,
+  "platform": "Windows-11-10.0.26200-SP0",
+  "benchmark_order": [
+    "bf16",
+    "int8"
+  ],
+  "notes": "Windows WDDM desktop GPU; one sequential suite per precision, two seeds per case; no multi-run statistical confidence intervals."
+}

evaluation/int8/environment.json

@@ -0,0 +1,149 @@
+{
+  "python": "3.13.5",
+  "platform": "Windows-11-10.0.26200-SP0",
+  "gpu": "NVIDIA GeForce RTX 5090",
+  "cuda": "12.8",
+  "packages": {
+    "torch": "2.10.0+cu128",
+    "torchvision": "0.25.0+cu128",
+    "diffusers": "0.41.0.dev0",
+    "transformers": "5.17.0",
+    "accelerate": "1.15.0",
+    "bitsandbytes": "0.50.2",
+    "safetensors": "0.8.0",
+    "huggingface-hub": "1.32.0",
+    "tokenizers": "0.23.2",
+    "numpy": "2.2.6",
+    "pillow": "12.2.0",
+    "psutil": "7.0.0"
+  },
+  "model_identity": {
+    "kind": "int8",
+    "base_revision": "b3179ad355be050328e483a9dfdd9e60cd62adfa",
+    "files": [
+      {
+        "path": "model_index.json",
+        "size": 447,
+        "sha256": "cf1ecd104ea090855d60d8cec0895c1e9d6ee41b2f87e231b678beecd1cf7809"
+      },
+      {
+        "path": "processor/added_tokens.json",
+        "size": 707,
+        "sha256": "c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680"
+      },
+      {
+        "path": "processor/chat_template.jinja",
+        "size": 5292,
+        "sha256": "3636d0f0bd6bef02654cdffdc447b79cb2cef8ab02cc75267345946291a489e4"
+      },
+      {
+        "path": "processor/merges.txt",
+        "size": 1671853,
+        "sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
+      },
+      {
+        "path": "processor/preprocessor_config.json",
+        "size": 782,
+        "sha256": "93585062a80db5e8ca038efc7726a3e6411d9db948472d81d63c6303993be8c5"
+      },
+      {
+        "path": "processor/special_tokens_map.json",
+        "size": 613,
+        "sha256": "76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd"
+      },
+      {
+        "path": "processor/tokenizer.json",
+        "size": 11422654,
+        "sha256": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4"
+      },
+      {
+        "path": "processor/tokenizer_config.json",
+        "size": 5445,
+        "sha256": "81ec7bb9530159b326c0bef1d0b6c33d392090524014ea3f0123a3c1eb9c2af5"
+      },
+      {
+        "path": "processor/video_preprocessor_config.json",
+        "size": 817,
+        "sha256": "59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1"
+      },
+      {
+        "path": "processor/vocab.json",
+        "size": 2776833,
+        "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
+      },
+      {
+        "path": "scheduler/scheduler_config.json",
+        "size": 485,
+        "sha256": "5895f3a167c14a967fe9ac70c64924ae5acc79799e0679fd12907e594a713cd1"
+      },
+      {
+        "path": "text_encoder/config.json",
+        "size": 2487,
+        "sha256": "da4f7a17ffbcaef0d09f9c31cab01d77e2c25905676818279254b5a8bb371150"
+      },
+      {
+        "path": "text_encoder/generation_config.json",
+        "size": 467,
+        "sha256": "7dc6dd5e325f76cd2f75d428d41049e1f8ac0c4775d0800c0885b88bfe3f4c7f"
+      },
+      {
+        "path": "text_encoder/model-00001-of-00003.safetensors",
+        "size": 4999822155,
+        "sha256": "9b8ae3bd2a1e4884abc48c659212427bbbae06fb1bc17e7173583704fc0f6bdd"
+      },
+      {
+        "path": "text_encoder/model-00002-of-00003.safetensors",
+        "size": 4996432636,
+        "sha256": "d28d1a8b3994a43582acc5e2da15ff38b632fae1466f41bc7710d27ec170aa9e"
+      },
+      {
+        "path": "text_encoder/model-00003-of-00003.safetensors",
+        "size": 27729689,
+        "sha256": "5896ffa9e2b2a52f9d496755fde050306f21c449bd66c2526b40979033c6e244"
+      },
+      {
+        "path": "text_encoder/model.safetensors.index.json",
+        "size": 139297,
+        "sha256": "d3fc335d23047182f40d458efec86a4b187a766e0d1639426b6a9b7ab5b1ab8a"
+      },
+      {
+        "path": "transformer/config.json",
+        "size": 1297,
+        "sha256": "1a941b5389b5985cd16314641d708a1333a65748bcd3f74b1b8b1cd74b828202"
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model-00001-of-00002.safetensors",
+        "size": 4989447305,
+        "sha256": "0d77daaa3129860f026567e0178fec4ed776f3cf2463da76a0823b9c381c3468"
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model-00002-of-00002.safetensors",
+        "size": 2267331623,
+        "sha256": "5e1b8ee3917833687adb662830438dc612ace2cf4bff7b8898515184a888fde1"
+      },
+      {
+        "path": "transformer/diffusion_pytorch_model.safetensors.index.json",
+        "size": 77984,
+        "sha256": "96963ae89098cbfdb83fb09bfe0b91e34d8dfe812c65da5e1fbe452878a79ebe"
+      },
+      {
+        "path": "vae/config.json",
+        "size": 2079,
+        "sha256": "9785d527b278cb8b210e9f48a8028d92a4190968ce7e24c8fc85d9d1b82f6ba2"
+      },
+      {
+        "path": "vae/diffusion_pytorch_model.safetensors",
+        "size": 1350989512,
+        "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+      }
+    ],
+    "fingerprint": "df2eb85e5f936d6b172d71605502055db9be34cb8590447488ce60f7e9d8dc22"
+  },
+  "model": "D:\\Git\\Qwen-Image-2.1\\models\\int8",
+  "load_seconds": 15.678335699999934,
+  "load_peak_allocated_bytes": 17281157120,
+  "offload": "model",
+  "warmup": true,
+  "generator_device": "cpu",
+  "cases_sha256": "2663ee3d489e1eb136e02b083963b8622146fc65ac301d6acc5a6aab21e63356"
+}

evaluation/int8/records.jsonl

@@ -0,0 +1,14 @@
+{"case_id": "portrait", "category": "portrait", "prompt": "A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.65371929999992, "peak_allocated_bytes": 24547198464, "peak_reserved_bytes": 24595398656, "process_rss_after_bytes": 24881610752, "image": "portrait-s42.png", "image_sha256": "e37187a2530d0f45378f17ca8fff47a7e174f1ce0c19ff89b9df8f4db1845acd", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 249, "alpha_max": 255, "alpha_nonopaque_fraction": 0.015732765197753906}
+{"case_id": "portrait", "category": "portrait", "prompt": "A natural documentary portrait of an elderly woman with silver hair and freckles, wearing a blue wool sweater, standing beside a window in soft morning light. Realistic skin texture, gentle expression, waist-up composition.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.614630900000066, "peak_allocated_bytes": 24547197952, "peak_reserved_bytes": 24595398656, "process_rss_after_bytes": 25830686720, "image": "portrait-s123.png", "image_sha256": "09391672b7aa15abc92a49184c94cd7dc09b570b2134c95f0a1524701630323c", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 252, "alpha_max": 255, "alpha_nonopaque_fraction": 0.08091926574707031}
+{"case_id": "english_text", "category": "typography_en", "prompt": "A clean editorial poster with a deep blue background. Large exact headline at the top: \"CREATE WITH LIGHT\". Smaller exact subtitle: \"September 2026\". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.416275399999904, "peak_allocated_bytes": 24547739136, "peak_reserved_bytes": 24601690112, "process_rss_after_bytes": 26035294208, "image": "english_text-s42.png", "image_sha256": "98ea66acd842498e192bc69503e32f5f61a4e5e56887d0f86e9c936b0b0d976f", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 250, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0012531280517578125}
+{"case_id": "english_text", "category": "typography_en", "prompt": "A clean editorial poster with a deep blue background. Large exact headline at the top: \"CREATE WITH LIGHT\". Smaller exact subtitle: \"September 2026\". A realistic orange desk lamp occupies the lower half. Elegant balanced typography, no additional text.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.404967399999805, "peak_allocated_bytes": 24547762688, "peak_reserved_bytes": 24601690112, "process_rss_after_bytes": 26170896384, "image": "english_text-s123.png", "image_sha256": "b55551e7c17975ee90ad895e449379355886c3422d6dcdee9b0c3a4ce1c4632f", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 213, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0015974044799804688}
+{"case_id": "chinese_text", "category": "typography_zh", "prompt": "设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.544581299999663, "peak_allocated_bytes": 24548032000, "peak_reserved_bytes": 24603787264, "process_rss_after_bytes": 26308521984, "image": "chinese_text-s42.png", "image_sha256": "dce2de1e3b8b3836c7584854c7cf340cf05bcb141fa8e3988264c62e1b20238c", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 247, "alpha_max": 255, "alpha_nonopaque_fraction": 0.01668834686279297}
+{"case_id": "chinese_text", "category": "typography_zh", "prompt": "设计一张现代咖啡馆海报，米白背景，中央一杯拉花拿铁。顶部准确写上中文大标题“慢下来，喝杯咖啡”，下方准确写上“小店今日营业”。文字清晰，留白充分，暖色摄影，不添加其他文字。", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.86057179999989, "peak_allocated_bytes": 24548050432, "peak_reserved_bytes": 24603787264, "process_rss_after_bytes": 26289606656, "image": "chinese_text-s123.png", "image_sha256": "7ff8508a9454650920591c0aad34d1793501f609e853f2f8584c15eee5127479", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 245, "alpha_max": 255, "alpha_nonopaque_fraction": 0.012395858764648438}
+{"case_id": "composition", "category": "spatial_counting", "prompt": "A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.130140400000073, "peak_allocated_bytes": 24547978240, "peak_reserved_bytes": 24603787264, "process_rss_after_bytes": 26264928256, "image": "composition-s42.png", "image_sha256": "5704252c99e7f9dbe9e512eb82bf4fd93650e26c1d6479412679e7868637bcdd", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 252, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0036773681640625}
+{"case_id": "composition", "category": "spatial_counting", "prompt": "A studio photograph on a light gray tabletop: exactly three ceramic cups in a row, a red cup on the left, a blue cup in the middle, and a yellow cup on the right. A single green apple sits in front of the blue cup. Soft shadows, no text.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.303240899999764, "peak_allocated_bytes": 24547941888, "peak_reserved_bytes": 24603787264, "process_rss_after_bytes": 27820789760, "image": "composition-s123.png", "image_sha256": "c030c007a9c3962533ffeb256d9dac619cc082a20a21ff6f9eb0c0ab7b7cadae", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 253, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0106048583984375}
+{"case_id": "texture", "category": "texture", "prompt": "Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.223099399999683, "peak_allocated_bytes": 24546856448, "peak_reserved_bytes": 24597495808, "process_rss_after_bytes": 28058947584, "image": "texture-s42.png", "image_sha256": "37042d1659147a142e0562aa27623cb1d159d9d12e46af59395029f0293fd9c8", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 251, "alpha_max": 255, "alpha_nonopaque_fraction": 0.0345611572265625}
+{"case_id": "texture", "category": "texture", "prompt": "Macro photography of a small kingfisher perched on a mossy branch beside clear water, detailed blue feathers, droplets, natural sunlight, softly blurred forest background, realistic textures.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.567528400000356, "peak_allocated_bytes": 24546861568, "peak_reserved_bytes": 24597495808, "process_rss_after_bytes": 28048392192, "image": "texture-s123.png", "image_sha256": "82fc88d5ab3cd4201634c54567b327f200231321dd0207990ff4c893e78496c7", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 251, "alpha_max": 255, "alpha_nonopaque_fraction": 0.01710224151611328}
+{"case_id": "rgba", "category": "transparency", "prompt": "This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.388802900000428, "peak_allocated_bytes": 24546763776, "peak_reserved_bytes": 24595398656, "process_rss_after_bytes": 28067987456, "image": "rgba-s42.png", "image_sha256": "8624d495fe40f602a492b3366ccb0354d5684e4e084281503fdb435d0f4e265e", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 0, "alpha_max": 255, "alpha_nonopaque_fraction": 0.7706136703491211}
+{"case_id": "rgba", "category": "transparency", "prompt": "This is an RGBA image with transparency. A cute cartoon dragon sticker, full body, green scales and small orange wings. The image has alpha channel and the background is transparent.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": null, "seconds": 19.288307799999984, "peak_allocated_bytes": 24546768384, "peak_reserved_bytes": 24595398656, "process_rss_after_bytes": 28076621824, "image": "rgba-s123.png", "image_sha256": "4871aab8f9d404a479385053bdb4c18de7f4119884bb0e30997d7be721a69b45", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 0, "alpha_max": 255, "alpha_nonopaque_fraction": 0.6456260681152344}
+{"case_id": "edit", "category": "image_editing", "prompt": "Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.", "seed": 42, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa", "seconds": 25.399254500000097, "peak_allocated_bytes": 24781434880, "peak_reserved_bytes": 26440892416, "process_rss_after_bytes": 29161738240, "image": "edit-s42.png", "image_sha256": "79725e243ccfd82b5387beffe6e1f37193ce9137b576b2ec20e8543547dd9d33", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 241, "alpha_max": 255, "alpha_nonopaque_fraction": 0.11248302459716797}
+{"case_id": "edit", "category": "image_editing", "prompt": "Change only the blue sweater to a red sweater. Preserve the same person, face, pose, lighting and background.", "seed": 123, "width": 1024, "height": 1024, "steps": 40, "cfg": 1.0, "kv_cache": true, "offload": "model", "input_sha256": "01e37e178fb0d35d5c706b7edd76bc1a6c5b50ec3536901a467fd45295d097aa", "seconds": 25.24017350000031, "peak_allocated_bytes": 24782129152, "peak_reserved_bytes": 26440892416, "process_rss_after_bytes": 29135400960, "image": "edit-s123.png", "image_sha256": "d8023e34ba10a22b36f63d050c50d44f149cebc7fc8f7f612a47f27f2a4cb70b", "image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 253, "alpha_max": 255, "alpha_nonopaque_fraction": 0.011339187622070312}

evaluation/qualitative.md

@@ -0,0 +1,24 @@
+# Qualitative inspection / 定性检查
+
+Agent visual inspection of the seed-42 pairs; not a blinded human preference study,
+OCR benchmark, or quantitative semantic score. All seed-123 outputs remain available
+in the gallery for independent review.
+
+| Case | Observed BF16 / INT8 behavior |
+|---|---|
+| Portrait | Both preserve the requested older woman, blue sweater and window composition. Face shape, wrinkles, framing and fabric details shift slightly. Neither is pixel-identical. |
+| English lettering | Both visibly render “CREATE WITH LIGHT” and “September 2026”. Layout and desk objects differ. |
+| Chinese lettering | Both visibly render the requested main title and subtitle. Both also add unwanted, partly nonsensical footer text despite the instruction to add no other text. |
+| Counting / arrangement | Both show three cups in red/blue/yellow order with one green apple in front of the blue cup. Object scales and spacing differ. |
+| Texture | Both depict a plausible blue/orange kingfisher. Head angle, branch orientation and feather texture differ noticeably. Low pixel similarity here is not by itself evidence of poor semantic quality. |
+| RGBA | Both produce a green/orange dragon sticker and a real varying alpha channel. Sampled background pixels have alpha 0–1/255; the subject is opaque. Invisible RGB background colors should not be mistaken for an opaque background. Both have a white sticker outline. |
+| Editing | Both recolor the sweater toward a red/burgundy tone, but produce conspicuous oversharpening and contrast/texture changes across the portrait and background. Faithful preservation of lighting and skin texture is not achieved in this test. |
+
+**Editing limitation:** The artifact is reproducible in both precisions under this
+specific pinned runtime and test setup. This comparison does not identify its root
+cause and does not attribute it specifically to INT8. Do not treat successful pipeline
+execution as validation of editing fidelity.
+
+中文要点：文生图整体语义在所检查样例中保留，但画面细节会变化。中英文主标题可读；
+中文海报的多余页脚在两种精度中均存在。编辑任务在两组都出现明显过度锐化及对比度变化，
+不能宣称“编辑无损”或“质量完全一致”。这些观察只针对所列样例，不代表通用榜单结论。

evaluation/report.md

@@ -0,0 +1,40 @@
+# BF16 / INT8 comparison
+
+Additional adult Chinese woman portrait pairs are reported separately: [supplement](young_woman/report.md).
+
+14 paired outputs on NVIDIA GeForce RTX 5090; 1024×1024, 40 steps, offload=model, CFG=1, KV cache enabled.
+
+Warmup excluded. Pixel metrics measure drift, not semantic quality. The same BF16 portrait is used as input for every editing pair. The suite is small and does not establish a general quality ranking.
+
+| Case | Seed | BF16 s | INT8 s | BF16 peak GiB | INT8 peak GiB | RGB MAE (white) |
+|---|---:|---:|---:|---:|---:|---:|
+| portrait | 42 | 29.77 | 19.65 | 16.40 | 22.86 | 0.0134 |
+| portrait | 123 | 26.55 | 19.61 | 16.40 | 22.86 | 0.0163 |
+| english_text | 42 | 25.95 | 19.42 | 16.41 | 22.86 | 0.0263 |
+| english_text | 123 | 25.88 | 19.40 | 16.41 | 22.86 | 0.0118 |
+| chinese_text | 42 | 25.47 | 19.54 | 16.41 | 22.86 | 0.0597 |
+| chinese_text | 123 | 25.85 | 19.86 | 16.41 | 22.86 | 0.0246 |
+| composition | 42 | 25.38 | 19.13 | 16.41 | 22.86 | 0.0227 |
+| composition | 123 | 25.35 | 19.30 | 16.41 | 22.86 | 0.0196 |
+| texture | 42 | 25.85 | 19.22 | 16.40 | 22.86 | 0.0698 |
+| texture | 123 | 25.76 | 19.57 | 16.40 | 22.86 | 0.0168 |
+| rgba | 42 | 26.04 | 19.39 | 16.40 | 22.86 | 0.0183 |
+| rgba | 123 | 25.07 | 19.29 | 16.40 | 22.86 | 0.0465 |
+| edit | 42 | 31.16 | 25.40 | 19.08 | 23.08 | 0.0097 |
+| edit | 123 | 31.03 | 25.24 | 19.08 | 23.08 | 0.0039 |
+
+## Summary
+
+Mean latency: BF16 26.79s; INT8 20.29s.
+
+[Interactive-sized side-by-side gallery](comparison.html). Raw data: comparison.csv, bf16/records.jsonl, int8/records.jsonl. Environment records include package versions and loading overhead.
+
+[Qualitative observations](qualitative.md).
+
+## Interpretation limits
+
+- No CLIP, OCR, human preference, FID or benchmark leaderboard score is claimed.
+- Peak CUDA allocated/reserved memory excludes other processes and display usage.
+- Measured call latency includes transfers; disk writing and model loading are excluded.
+- Paired images may diverge with quantization even when both remain plausible.
+- Editing and transparency should be inspected in the gallery, including preserved details.