Release verified Image21-INT8 conversion

.gitattributes

@@ -122,3 +122,108 @@ evaluation/editing/inputs/portrait_recolor.png filter=lfs diff=lfs merge=lfs -te
 evaluation/editing/int8/bird_background-s42.png filter=lfs diff=lfs merge=lfs -text
 evaluation/editing/int8/cup_material-s42.png filter=lfs diff=lfs merge=lfs -text
 evaluation/editing/int8/portrait_recolor-s42.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-bf16-full/decoded-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-bf16-full/decoded-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-bf16-full/decoded-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-bf16-full/decoded-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-bf16-full/decoded.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-fp32-tiled/decoded-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-fp32-tiled/decoded-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-fp32-tiled/decoded-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-fp32-tiled/decoded-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decode-fp32-tiled/decoded.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/decoder-crops.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/int8-crops.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/latent-2048/bird_background-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/latent-2048/diagnostics/bird_background-s1000042-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/latent-2048/diagnostics/bird_background-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/latent-2048/diagnostics/bird_background-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/latent-2048/diagnostics/bird_background-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/paired-birds.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/probe/bird_background-1024-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/probe/bird_background-2048-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/audit/probe/cup_material-2048-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/bird_background-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/bird_background-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/bird_remove_background-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/bird_remove_background-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/cup_identity-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/cup_identity-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/cup_material-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/cup_material-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_background-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000042-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000123-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/bird_remove_background-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_identity-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/cup_material-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/diagnostics/portrait_recolor-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/portrait_recolor-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/bf16/portrait_recolor-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/inputs/bird_background.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/inputs/cup_material.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/inputs/portrait_recolor.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/bird_background-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/bird_background-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/bird_remove_background-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/bird_remove_background-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/cup_identity-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/cup_identity-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/cup_material-s1000042.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/cup_material-s1000123.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_background-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000042-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000123-alpha.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/bird_remove_background-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_identity-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/cup_material-s1000123-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/portrait_recolor-s1000042-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/portrait_recolor-s1000042-checker.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/portrait_recolor-s1000042-white.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/portrait_recolor-s1000123-black.png filter=lfs diff=lfs merge=lfs -text
+evaluation/editing-v2/int8/diagnostics/portrait_recolor-s1000123-checker.png filter=lfs diff=lfs merge=lfs -text

EDITING_UPDATE.md

@@ -0,0 +1,45 @@
+## Editing evaluation revision — 2026-09-24
+
+**Weights are unchanged.** This update corrects the editing evaluation protocol and inference defaults.
+Historical images, measurements and measured scripts remain available for reproduction.
+
+The old seed-42 edits reused the source-generation seed. When the RNG stream and latent shape also
+match, this can replay the source noise and confound editing quality. See the corrected discussion in
+[Diffusers #14824](https://github.com/huggingface/diffusers/issues/14824).
+Changing seed alone did **not** remove the artifacts in our 2048px tiled controls; the previous general
+2048px recommendation is withdrawn.
+
+**The short vertical stains were isolated to tiled VAE decoding in the audited 2048px example.**
+Decoding the exact same saved latent with the same VAE checkpoint without tiling removes the
+visible stains; FP32 decoding with tiling retains them. This does not repair the missing branch
+or coarse bird detail already present in that failed edit. The quality suite therefore disables tiling.
+
+The revised suite uses five cases × two independent editing seeds (1000042 / 1000123), paired BF16
+and INT8, 1024px area scale, 40 steps, CFG=1, KV cache, model CPU offload and untiled VAE.
+It adds actual transparent-background removal and an unchanged-image control. The complete matrix,
+raw RGBA files, alpha maps, white/black/checker composites, hashes, schedules and decoder latents are
+in [evaluation/editing-v2](evaluation/editing-v2). Inspect the
+[visual audit](evaluation/editing-v2/AUDIT.md) and download
+[comparison.html](evaluation/editing-v2/comparison.html) with its image folders for local viewing.
+
+No filtering, sharpening or alpha cleanup was applied. These are informal samples, not an official
+benchmark or a guarantee that all artifacts are resolved. Latent capture was enabled for diagnosis,
+so this revision makes no latency/speedup claim.
+
+In the revised samples, snowy-background replacement retains the bird and branch, and background
+removal produces real transparency. Preservation is still imperfect: in the unchanged-image control,
+seed 1000123 changes the green apple to yellow in both precisions. This failed control is retained.
+
+Start editing with a seed independent of the source:
+
+```bash
+python -m scripts.infer --model . --input input.png --source-seed 42 --seed 1000042 --resolution 1024 --prompt "Change only the blue sweater to red. Preserve the person, lighting and background." --output edit.png
+```
+
+Omit `--source-seed` if unknown. Explicit seeds are never silently remapped; a known collision is
+rejected. Omitting width/height preserves the last input's aspect ratio. `--resolution` controls
+reference image area, while an explicit `--width` / `--height` pair controls output dimensions.
+VAE tiling remains an optional memory tradeoff, not a quality improvement guarantee.
+
+The sections below retain the historical v1 evaluation; their image-generation timings are not
+pooled with this revised editing suite.

benchmarks/editing-v2.json

@@ -0,0 +1,52 @@
+[
+  {
+    "id": "portrait_recolor",
+    "category": "clothing color",
+    "prompt": "Change only the cream sweater to a navy blue sweater. Preserve the same woman, face, hairstyle, pose, lighting and background.",
+    "input_file": "portrait_recolor.png",
+    "input_sha256": "00716ad4721f585be4b0bf0daf2413f4ab92b795f1eb4d08b4f1ecc604b075b8",
+    "source": "evaluation/young_woman/bf16/young_chinese_woman-s42.png",
+    "source_seed": 42,
+    "task": "opaque_edit"
+  },
+  {
+    "id": "cup_material",
+    "category": "material replacement",
+    "prompt": "Change only the middle blue ceramic cup into a brushed silver metal cup. Preserve its shape and position, the red and yellow cups, the green apple, the tabletop, lighting and background.",
+    "input_file": "cup_material.png",
+    "input_sha256": "91e967685fbbd79ba13daece22a0f72c863cb518db88d5d2a6f9aecfc5f292b1",
+    "source": "evaluation/bf16/composition-s42.png",
+    "source_seed": 42,
+    "task": "opaque_edit"
+  },
+  {
+    "id": "bird_background",
+    "category": "background replacement",
+    "prompt": "Replace only the blurred forest background with a softly blurred snowy winter forest. Preserve the same kingfisher, its colors and pose, the mossy branch, framing and foreground lighting.",
+    "input_file": "bird_background.png",
+    "input_sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095",
+    "source": "evaluation/bf16/texture-s42.png",
+    "source_seed": 42,
+    "task": "opaque_edit"
+  },
+  {
+    "id": "bird_remove_background",
+    "category": "background removal",
+    "prompt": "Remove the background from <image1>. Keep the entire kingfisher and the mossy branch unchanged, including fine feather edges and feet. This is an RGBA image with transparency. The image has alpha channel and the background is transparent.",
+    "input_file": "bird_background.png",
+    "input_sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095",
+    "source": "evaluation/bf16/texture-s42.png",
+    "source_seed": 42,
+    "task": "alpha_extraction"
+  },
+  {
+    "id": "cup_identity",
+    "category": "identity control",
+    "prompt": "Keep <image1> unchanged. Preserve all objects, colors, details, lighting and background exactly.",
+    "input_file": "cup_material.png",
+    "input_sha256": "91e967685fbbd79ba13daece22a0f72c863cb518db88d5d2a6f9aecfc5f292b1",
+    "source": "evaluation/bf16/composition-s42.png",
+    "source_seed": 42,
+    "task": "identity"
+  }
+]

cards/huggingface.md

@@ -65,8 +65,9 @@ image.save("output.png")
 ```
 
 The saved quantization configuration is detected automatically. Do not re-quantize
-these weights at load time. For editing, use the 2048px editing recipe below, setting
-`width`, `height` and `output_resolution` together and enabling VAE tiling.
+these weights at load time. For editing, use a seed independent of the source image
+(for example 1000042 when the source used 42), starting at output_resolution=1024.
+The retained 2048px section is historical evidence, not a general recommendation.
 For transparency, use the upstream RGBA prompt format and save PNG files.
 
 Use the bundled loader on the pinned runtime: it loads/offloads the two INT8
@@ -76,6 +77,10 @@ this tested stack. The helper does not change the checkpoint or quantization.
 
 ## Informal release evaluation
 
+Historical v1 editing reused the source seed 42. This confounds the 1024px editing
+quality assessment through possible noise replay. See `EDITING_UPDATE.md` and
+`evaluation/editing-v2/` when present for the revised independent-seed evaluation.
+
 **This community evaluation accompanies this model release and is provided for reference only. It does not represent any official evaluation by Qwen, Hugging Face, ModelScope, or their teams.** Results reflect a small, locally selected sample on one hardware/software configuration; they are not a standardized ranking or a guarantee of quality, speed, or memory requirements.
 
 {{EVALUATION_EN}}

cards/modelscope.md

@@ -65,8 +65,8 @@ 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。
-图像编辑请使用下文 2048px 编辑代码，同时设置 `width`、`height`、
-`output_resolution`，并开启 VAE 分块。
+图像编辑请使用与源图独立的种子（如源图为 42 时编辑用 1000042），
+从 `output_resolution=1024` 起步。下文 2048px 部分为历史证据，不是通用推荐。
 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。
 
 在固定版本环境下请使用仓库附带的加载函数：它依次加载并卸载两个 INT8 组件，

docs/editing-audit-20260924.md

@@ -0,0 +1,83 @@
+# 改图评测审计（2026-09-24）
+
+本次更新只修订推理入口、评测与说明，不修改、重新量化或重新下载模型权重。
+历史评测文件及其测量实现保留。新协议名称为 `editing-v2`。
+
+## 已确认的评测问题
+
+1. 旧图像来源为 BF16、CPU generator、seed 42；旧编辑同样固定 seed 42。
+   相同随机流、种子和噪声张量形状会复用生成源图的初始噪声。
+   上游 issue #14824 的报告者确认换编辑种子后，其过锐化/光晕失效消失。
+   这使旧 1024px seed-42 编辑不能作为无混淆的质量结论。
+   2048px 改变了噪声形状，不能把它的所有瑕疵也都归因于精确噪声重放。
+2. 旧 2048px 结果同时改变输出面积、参考图缩放、调度器序列长度对应的 sigma 和
+   VAE 分块，单个毛衣样例改善不足以支持统一推荐 2048px。
+3. `output_resolution=width` 混淆面积尺度和横向像素数；非方形图固定输出尺寸还会
+   改变构图。新入口分开设置面积尺度与显式输出尺寸，并记录实际尺寸。
+4. 背景替换不等于背景移除。旧雪林输出几乎不透明，不能用于透明边缘评价。
+5. 旧编辑组只用一个种子，缺少独立背景移除、保持原图控制及 alpha/多底色检查。
+
+## 核对过的实现和参数
+
+- 固定 Diffusers 提交 `80c7ed262aeffbeb43ef13ae04baeb9b84515a69`。
+  2026-09-24 下载的官方 main 提交 `e0118ade2f60234c41bacf40330a7e2f61108849` 中，
+  QwenImage21 pipeline 与 transformer 源码与本地逐行相同；没有证据支持盲目升级依赖。
+- 40 步、CFG=1、KV cache 是本次固定对照设置。未随意增加 CFG 或添加未生效的负面词。
+- 输入 RGBA 保留；上游按训练约定只对白底合成视觉编码器副本，VAE 接收 RGBA。
+  不将 `.convert('RGB')` 作为抠图修复；不对输出阈值化、腐蚀、平滑或锐化。
+- CPU 随机数生成器、模型 CPU 卸载、INT8 辅助张量修复均保持不变。
+- VAE 分块是内存优化，必须与不分块对照区分。FP32 解码实验只改变解码器计算精度，
+  不会恢复已经在 BF16 编码/去噪阶段损失的信息。
+
+## 修订协议
+
+共同输入及其 SHA256 沿用历史评测。固定编辑种子 1000042/1000123，源图种子声明为 42；
+发现种子冲突或重复、输入哈希变化、输出目录已存在则拒绝运行。
+默认 1024 面积尺度、40 步、CFG=1、KV cache、不分块；BF16/INT8 在独立进程运行。
+
+五类用例：人物衣服改色、杯子改材质、鸟图背景替换、鸟与枝干透明背景提取、杯子保持原图。
+背景替换沿用原提示词；背景移除新增显式 RGBA 指令。保持原图结果是误改诊断，
+不要求生成式编辑器逐像素无损返回输入。
+
+每张保留原始 PNG、实际模式/尺寸、输入/输出哈希、真实 sigma 序列和 alpha 分布，
+另存 alpha 图、黑底、白底、棋盘底，供 100% 放大检查羽毛/杯沿及平坦背景。
+alpha≤5、alpha≥250 分别作为透明/近不透明的描述阈值，不作为成功判定。
+没有真实 alpha 标注或目标区域 mask，故不宣称 matting SAD/边缘 F1/身份保持分数。
+BF16/INT8 像素差异仅描述两种精度的偏移，不能替代编辑成功率。
+
+本次审计开启 `--save-latents`，其保存开销使记录耗时不用于性能排名。
+正常性能重测应去掉该开关，保留完整预热，并保持进程隔离。
+
+## 单变量对照（已观察）
+
+**短竖污损已定位到默认 VAE 分块解码路径。** 对同一个 2048px 去噪后潜变量：
+BF16 分块有明显短竖痕迹；仅关闭分块后所检查背景的痕迹消失；FP32 保持分块仍有痕迹。
+三个实验使用同一份 VAE checkpoint，未改变权重文件。重生成的故障 PNG 与初次探针逐字节一致。
+这证明本例竖纹由分块解码引入，但不等于解决了同图的语义编辑失败。
+不分块解码仍保留枝干丢失和鸟体粗糙细节，不能靠解码修复其上游编码/去噪结果。
+
+固定 INT8、旧提示词、40 步、CFG=1、分块：
+
+- 2048px 杯子从 seed 42 改为 1000042 后仍有细竖纹；换种子不等于解决全部污损。
+- 2048px 鸟图改用 1000042 后仍出现白底、枝干丢失和粗糙细节。
+- 同种子 1000042 改为 1024px 后，鸟与枝干得以保留，背景出现变化；这并不说明雪林
+  指令已被完整准确执行，也不能推广为所有素材的最佳分辨率。
+
+完整新协议和同潜变量 VAE 对照结果以发布包的 `evaluation/editing-v2/AUDIT.md` 为准。
+不能由当前证据断言所有剩余瑕疵均非权重/训练分布所致。
+
+## 复现及发布准备
+
+运行命令见根 README 的“修订改图评测（v2）”。
+`scripts.editing_report` 会拒绝不完整矩阵、种子/参数/哈希不一致或伪造透明度记录。
+`scripts.release` 对新评测核对模型身份、实测源码哈希和配对调度轨迹。
+平台候选��保留历史数据，将 v2 放入独立目录并在模型卡顶部解释方法修正。
+上传只做 `--dry-run`；实际发布仍使用经过校验的完整候选目录。
+
+## 上游依据
+
+- https://github.com/huggingface/diffusers/issues/14824 （包含后续更正，不能只读最初的故障描述）
+- https://github.com/huggingface/diffusers/tree/e0118ade2f60234c41bacf40330a7e2f61108849
+- https://github.com/QwenLM/Qwen-Image-2.1
+
+下载快照、issue 评论与提交信息保存在 `artifacts/upstream-audit-20260924/`。

evaluation/editing-v2/AUDIT.md

@@ -0,0 +1,78 @@
+# Editing v2 visual audit / 改图 v2 逐图审计
+
+2026-09-24. Five cases, two seeds (1000042 and 1000123), BF16 and INT8: **20 outputs / 10 pairs**.
+Weights, fixed dependency versions and CPU offload implementation are unchanged.
+All measured runs completed with exit code 0. `process-isolation.jsonl` records launcher and actual
+inference PIDs separately (Windows venv uses a launcher); original launcher records are retained
+in `process-launchers-original.jsonl`. Actual PIDs are taken from each measured `environment.json`.
+
+Parameters: 1024px area scale, 40 steps, CFG=1, KV cache enabled, untiled BF16 VAE, CPU generator.
+All inputs are 1024px BF16-generated RGBA images with source seed 42, shared by both precisions.
+Source seeds never equal edit seeds. Full warmup is excluded, but diagnostic latent saving adds
+overhead, so times are **not** used for speed or latency claims.
+
+## Observations / 观察
+
+Every case/seed pair was inspected. Original PNGs and alpha/white/black/checker displays are linked
+in `comparison.html`. Enlarged crops are display aids only; no scored output was retouched.
+
+| Case | Both seeds, both precisions / 两种精度与两个种子的观察 |
+|---|---|
+| portrait_recolor | Cream sweater becomes navy. Broad face, pose and window lighting remain, but facial/hair/fabric details drift. / 毛衣改为藏蓝，大体保留人物与光线，五官、发丝、织物并非无损保持。 |
+| cup_material | Middle cup becomes silver metal, while three-cup ordering and green apple remain. Inspected flat-background crops no longer show the conspicuous short vertical stains of the historical 2048px samples; subtle texture remains. / 中杯变为银色金属，排列与绿苹果保留；抽查平坦背景未见旧 2048px 样例的明显短竖污损，仍有细纹理。 |
+| bird_background | Snowy/white blurred forest appears while bird and mossy branch remain. Background structure and feather/branch detail change. / 出现雪色虚化林景，保留鸟和苔藓枝干，背景布局、羽毛与枝干细节仍变化。 |
+| bird_remove_background | Real transparent background; bird and branch retained. Inspected head/beak edges look substantially cleaner than the old 2048px white-background failure, but there is no ground-truth alpha mask to establish matting accuracy. / 真实透明背景，鸟与枝干保留，抽查头部与鸟喙边缘较旧白底失败样例规整；没有 alpha 真值，不能认定为精确抠图。 |
+| cup_identity | Seed 1000042 broadly retains the scene. **Seed 1000123 changes the green apple to yellow in BOTH BF16 and INT8 despite the unchanged-image instruction.** This control fails and is not hidden. / seed 1000042 大体保持原图；**1000123 在两种精度下都误改了苹果颜色**，保留该失败，不将这一组宣称为无损编辑。 |
+
+## Alpha inspection / 透明度检查
+
+The four background-removal outputs have roughly 71.2–71.8% near-transparent pixels (alpha≤5/255),
+and roughly 0.8% soft-alpha pixels (5<alpha<250). Exact values are in raw records.
+These descriptive fractions are **not** quality scores: transparent area does not prove correct edges.
+Opaque edit outputs have no near-transparent background in this suite.
+
+四张背景移除图透明像素占比约 71.2%–71.8%，过渡像素约 0.8%。统计仅用来区分真实透明与白底，
+不是质量评分；不能把本应保留的主体被删掉后得到的“大透明面积”当作成功。
+
+## Methodological limits / 方法限制
+
+- The historical 1024px seed-42 edit reused source noise; independent seeds remove this known confound.
+  Upstream explanation and corrections: https://github.com/huggingface/diffusers/issues/14824 .
+- Seed change alone does not fix the retained 2048px tiled failures. Lower resolution and VAE tiling
+  are independent factors and must not be collapsed into a quantization conclusion.
+- Source fixtures are generated by the same base model. This is a small regression set, not an
+  external-photograph benchmark, a human identity test, or a standardized editing ranking.
+- No target-region masks, ground-truth alpha mattes, blinded ratings, OCR/CLIP/FID scores or
+  claims of universal artifact removal. Paired pixel differences only describe BF16/INT8 drift.
+- Exact original source files for the historical main/editing suites remain shipped. New protocol
+  source hashes and upstream class hashes are recorded separately. No weights or scheduler configs
+  were changed to obtain these results.
+
+原始故障图片、单变量分辨率/种子对照与同潜变量解码实验保留在本目录 `audit/`。
+剩余失败不能一概排除模型权重、训练分布或生成式编辑本身的影响。
+
+## Same-latent VAE diagnosis / 同潜变量 VAE 解码定位
+
+The 2048px INT8 bird edit at seed 1000042 was reproduced in a fresh process with a PNG hash
+identical to the initial probe. Its denormalized decoder input is retained in
+`audit/latent-2048/bird_background-s1000042-decode.pt`.
+
+| Decoder | Observed flat-background stains |
+|---|---|
+| BF16, default 256px tile / 192px stride | Obvious short vertical stains |
+| BF16, tiling disabled | Stains absent in inspected matching background region |
+| FP32, same tiling enabled | Stains remain |
+
+Each decoder used the same saved latent hash and the same checkpoint bytes; no denoising was
+rerun for the alternate decoders. The full BF16 decode and tiled FP32 decode both completed.
+This isolates the visible short vertical stains in this sample to tiled VAE decoding, rather
+than INT8 quantization or BF16 decoder arithmetic alone. It does not prove every artifact in
+every image has the same cause. The missing branch and degraded bird texture persist after full
+decoding, so they cannot be repaired by merely switching decoder precision or tiling.
+
+三条解码路径使用同一潜变量和同一 checkpoint。短竖痕迹随分块开关变化，而不随单独提升
+解码精度消失，因此本例可定位为分块解码伪影。语义失败仍保留，不把所有问题都归入该结论。
+
+See `audit/decoder-compare-all.jpg`, `audit/decoder-crops.png` (unscaled matching crops),
+individual raw PNGs and `audit/decoder-summary.json` for hashes and pairwise pixel differences.
+No cleanup filter or postprocessing was applied to any raw output.

evaluation/editing-v2/audit/decode-bf16-full/record.json

@@ -0,0 +1,30 @@
+{
+  "dtype": "bf16",
+  "tiling": false,
+  "seconds": 29.58638380000002,
+  "latents_sha256": "19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246",
+  "image_sha256": "a32e5d14545ea684005e1373a0308330682d15a6159a0db7020b74cafc1cb610",
+  "vae_source_sha256": "afb341db5e9d081e568ae4703119d1141d124aaf2fa9c061b3d4eb80ce73c15b",
+  "vae_weights": [
+    {
+      "path": "diffusion_pytorch_model.safetensors",
+      "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+    }
+  ],
+  "image_mode": "RGBA",
+  "actual_size": [
+    2048,
+    2048
+  ],
+  "alpha_min": 248,
+  "alpha_max": 255,
+  "alpha_transparent_fraction": 0.0,
+  "alpha_opaque_fraction": 0.9999995231628418,
+  "alpha_soft_fraction": 4.76837158203125e-07,
+  "alpha_thresholds": [
+    5,
+    250
+  ],
+  "rgb_dx_mean": 0.020006714388728142,
+  "rgb_dy_mean": 0.019580036401748657
+}

evaluation/editing-v2/audit/decode-fp32-tiled/record.json

@@ -0,0 +1,30 @@
+{
+  "dtype": "fp32",
+  "tiling": true,
+  "seconds": 18.333071200000177,
+  "latents_sha256": "19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246",
+  "image_sha256": "be79dfd73714a8be00e812af65d81aee0269a883be2fe895ec46a1a487e968a2",
+  "vae_source_sha256": "afb341db5e9d081e568ae4703119d1141d124aaf2fa9c061b3d4eb80ce73c15b",
+  "vae_weights": [
+    {
+      "path": "diffusion_pytorch_model.safetensors",
+      "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+    }
+  ],
+  "image_mode": "RGBA",
+  "actual_size": [
+    2048,
+    2048
+  ],
+  "alpha_min": 194,
+  "alpha_max": 255,
+  "alpha_transparent_fraction": 0.0,
+  "alpha_opaque_fraction": 0.998399019241333,
+  "alpha_soft_fraction": 0.0016009807586669922,
+  "alpha_thresholds": [
+    5,
+    250
+  ],
+  "rgb_dx_mean": 0.018331428989768028,
+  "rgb_dy_mean": 0.018168220296502113
+}

evaluation/editing-v2/audit/decoder-processes.jsonl

@@ -0,0 +1,3 @@
+{"label": "latent-2048", "exit_code": 0, "args": ["-m", "scripts.benchmark_edits_v2", "--model", "models/int8", "--inputs", "artifacts\\editing-audit-20260924\\suite\\inputs", "--output", "artifacts\\editing-audit-20260924\\latent-2048", "--case", "bird_background", "--seeds", "1000042", "--resolution", "2048", "--vae-tiling", "--save-latents", "--no-warmup"]}
+{"label": "decode-bf16-full", "exit_code": 0, "args": ["-m", "scripts.audit_vae_decode", "--model", "models/bf16", "--latents", "artifacts\\editing-audit-20260924\\latent-2048\\bird_background-s1000042-decode.pt", "--output", "artifacts\\editing-audit-20260924\\decode-bf16-full", "--dtype", "bf16"]}
+{"label": "decode-fp32-tiled", "exit_code": 0, "args": ["-m", "scripts.audit_vae_decode", "--model", "models/bf16", "--latents", "artifacts\\editing-audit-20260924\\latent-2048\\bird_background-s1000042-decode.pt", "--output", "artifacts\\editing-audit-20260924\\decode-fp32-tiled", "--dtype", "fp32", "--tiling"]}

evaluation/editing-v2/audit/decoder-summary.json

@@ -0,0 +1,82 @@
+{
+  "decode-bf16-full": {
+    "dtype": "bf16",
+    "tiling": false,
+    "seconds": 29.58638380000002,
+    "latents_sha256": "19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246",
+    "image_sha256": "a32e5d14545ea684005e1373a0308330682d15a6159a0db7020b74cafc1cb610",
+    "vae_source_sha256": "afb341db5e9d081e568ae4703119d1141d124aaf2fa9c061b3d4eb80ce73c15b",
+    "vae_weights": [
+      {
+        "path": "diffusion_pytorch_model.safetensors",
+        "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+      }
+    ],
+    "image_mode": "RGBA",
+    "actual_size": [
+      2048,
+      2048
+    ],
+    "alpha_min": 248,
+    "alpha_max": 255,
+    "alpha_transparent_fraction": 0.0,
+    "alpha_opaque_fraction": 0.9999995231628418,
+    "alpha_soft_fraction": 4.76837158203125e-07,
+    "alpha_thresholds": [
+      5,
+      250
+    ],
+    "rgb_dx_mean": 0.020006714388728142,
+    "rgb_dy_mean": 0.019580036401748657,
+    "vs_bf16_tiled": {
+      "rgb_mae_white": 0.007886896841228008,
+      "rgb_psnr_white_db": 33.39505014952212,
+      "alpha_mae": 0.00024037367256823927
+    }
+  },
+  "decode-fp32-tiled": {
+    "dtype": "fp32",
+    "tiling": true,
+    "seconds": 18.333071200000177,
+    "latents_sha256": "19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246",
+    "image_sha256": "be79dfd73714a8be00e812af65d81aee0269a883be2fe895ec46a1a487e968a2",
+    "vae_source_sha256": "afb341db5e9d081e568ae4703119d1141d124aaf2fa9c061b3d4eb80ce73c15b",
+    "vae_weights": [
+      {
+        "path": "diffusion_pytorch_model.safetensors",
+        "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+      }
+    ],
+    "image_mode": "RGBA",
+    "actual_size": [
+      2048,
+      2048
+    ],
+    "alpha_min": 194,
+    "alpha_max": 255,
+    "alpha_transparent_fraction": 0.0,
+    "alpha_opaque_fraction": 0.998399019241333,
+    "alpha_soft_fraction": 0.0016009807586669922,
+    "alpha_thresholds": [
+      5,
+      250
+    ],
+    "rgb_dx_mean": 0.018331428989768028,
+    "rgb_dy_mean": 0.018168220296502113,
+    "vs_bf16_tiled": {
+      "rgb_mae_white": 0.0013088202103972435,
+      "rgb_psnr_white_db": 51.63176635617525,
+      "alpha_mae": 8.16673637018539e-05
+    }
+  },
+  "bf16_tiled": {
+    "image_sha256": "177fc897274faa4846659ec273d8ae39453d67526dc91d2bcc71407ce4ccd9da",
+    "repeat_png_identical": true,
+    "vae_weights": [
+      {
+        "path": "diffusion_pytorch_model.safetensors",
+        "sha256": "a07a1b7c4ee2966a1b3bdc37de9b4f983d56937e46619f709a80b6e490675417"
+      }
+    ]
+  }
+}

evaluation/editing-v2/audit/latent-2048/COMPLETE.json

@@ -0,0 +1 @@
+{"records": 1, "records_sha256": "92c93e44d324bf3201a04c7ade2b725670793e05603430fe4e8bd78c47e08e4a"}

evaluation/editing-v2/audit/latent-2048/bird_background-s1000042-decode.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246
+size 2099025

evaluation/editing-v2/audit/latent-2048/cases.json

@@ -0,0 +1,12 @@
+[
+  {
+    "id": "bird_background",
+    "category": "background replacement",
+    "prompt": "Replace only the blurred forest background with a softly blurred snowy winter forest. Preserve the same kingfisher, its colors and pose, the mossy branch, framing and foreground lighting.",
+    "input_file": "bird_background.png",
+    "input_sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095",
+    "source": "evaluation/bf16/texture-s42.png",
+    "source_seed": 42,
+    "task": "opaque_edit"
+  }
+]

evaluation/editing-v2/audit/latent-2048/environment.json

@@ -0,0 +1,199 @@
+{
+  "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"
+  },
+  "protocol": "editing-v2",
+  "pid": 13924,
+  "before_load_memory": {
+    "allocated_bytes": 0,
+    "reserved_bytes": 0,
+    "cuda_free_bytes": 32496418816,
+    "cuda_total_bytes": 34162016256
+  },
+  "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"
+  },
+  "offload": "model",
+  "offload_aux_fix": true,
+  "warmup": false,
+  "generator_device": "cpu",
+  "seeds": [
+    1000042
+  ],
+  "resolution": 2048,
+  "steps": 40,
+  "kv_cache": true,
+  "vae_tiling": true,
+  "cases_sha256": "bc5fae0b7a9e9102f34a8e56db4503c0c6d8903b5c873b67b37cdfa47438572a",
+  "source_sha256": {
+    "benchmark_edits_v2.py": "e03a4d623c382dc55183ff1beb4d7eb5ee5612ab6bb84341c45a8c08495313ad",
+    "editing_protocol.py": "a6e683f5bfba35fc19484567feb32a50edcde33ed06918fad63c23af1806d8c4",
+    "benchmark.py": "39544666c2e94a0d56de5057c7ab1e88df90dd5ba4d2cba2428c11f9d7c4d9fb",
+    "runtime.py": "26db92afa7d5c72324080480abe52c8e677b899844b369d99ac895e9bb1fb71c"
+  },
+  "scheduler_config": {
+    "num_train_timesteps": 1000,
+    "shift": 1.0,
+    "use_dynamic_shifting": true,
+    "base_shift": 0.5,
+    "max_shift": 0.9,
+    "base_image_seq_len": 256,
+    "max_image_seq_len": 8192,
+    "invert_sigmas": false,
+    "shift_terminal": 0.02,
+    "use_karras_sigmas": false,
+    "use_exponential_sigmas": false,
+    "use_beta_sigmas": false,
+    "time_shift_type": "exponential",
+    "stochastic_sampling": false,
+    "_class_name": "FlowMatchEulerDiscreteScheduler",
+    "_diffusers_version": "0.37.0.dev0"
+  },
+  "vae_dtype": "torch.bfloat16",
+  "upstream_source_sha256": {
+    "QwenImage21Pipeline": "6985b2f1f25e8dd09ef85c867ebcad184b981f10c0b4f80757ca9cc135b8fa5e",
+    "AutoencoderKLQwenImage21": "afb341db5e9d081e568ae4703119d1141d124aaf2fa9c061b3d4eb80ce73c15b",
+    "QwenImage21Transformer2DModel": "0eb0555e21ca93195e1fe9389113cafbf0e8822f6454c3001cebb3d21521ebd7",
+    "FlowMatchEulerDiscreteScheduler": "1af27be5b2f92b7d139d3c50239be5ce3eafc4a7eddf59c2d30689f8ec31e93d"
+  },
+  "vae_tiles": {
+    "tile_sample_min_height": 256,
+    "tile_sample_min_width": 256,
+    "tile_sample_stride_height": 192,
+    "tile_sample_stride_width": 192
+  }
+}

evaluation/editing-v2/audit/latent-2048/records.jsonl

@@ -0,0 +1 @@
+{"protocol": "editing-v2", "case_id": "bird_background", "category": "background replacement", "prompt": "Replace only the blurred forest background with a softly blurred snowy winter forest. Preserve the same kingfisher, its colors and pose, the mossy branch, framing and foreground lighting.", "seed": 1000042, "source_seed": 42, "input_file": "bird_background.png", "input_sha256": "40a0f73bb407a394fa9e8a3f6922455df716b9725ad07437fa10c1daca2c2095", "input_diagnostics": {"image_mode": "RGBA", "actual_size": [1024, 1024], "alpha_min": 251, "alpha_max": 255, "alpha_transparent_fraction": 0.0, "alpha_opaque_fraction": 1.0, "alpha_soft_fraction": 0.0, "alpha_thresholds": [5, 250], "rgb_dx_mean": 0.019595997408032417, "rgb_dy_mean": 0.019570155069231987}, "width": 2048, "height": 2048, "output_resolution": 2048, "steps": 40, "cfg": 1.0, "kv_cache": true, "vae_tiling": true, "offload": "model", "seconds": 202.94168359999958, "before_memory": {"allocated_bytes": 0, "reserved_bytes": 0, "cuda_free_bytes": 32496418816, "cuda_total_bytes": 34162016256}, "peak_allocated_bytes": 23254572544, "peak_reserved_bytes": 25222447104, "image": "bird_background-s1000042.png", "image_sha256": "177fc897274faa4846659ec273d8ae39453d67526dc91d2bcc71407ce4ccd9da", "sigmas": [1.0, 0.9926459789276123, 0.985012948513031, 0.9770848155021667, 0.9688441157341003, 0.9602718353271484, 0.9513477683067322, 0.9420496821403503, 0.9323535561561584, 0.9222333431243896, 0.9116604924201965, 0.9006039500236511, 0.8890299201011658, 0.8769010901451111, 0.8641765117645264, 0.850811243057251, 0.8367556929588318, 0.8219550251960754, 0.8063481450080872, 0.7898675203323364, 0.7724375128746033, 0.7539740204811096, 0.7343817949295044, 0.7135545015335083, 0.6913712024688721, 0.6676953434944153, 0.642370879650116, 0.6152193546295166, 0.5860357284545898, 0.5545830726623535, 0.5205857753753662, 0.4837222695350647, 0.44361412525177, 0.3998132348060608, 0.3517848253250122, 0.29888612031936646, 0.24033641815185547, 0.17517942190170288, 0.1022295355796814, 0.019999980926513672, 0.0], "latency_comparable": false, "image_mode": "RGBA", "actual_size": [2048, 2048], "alpha_min": 194, "alpha_max": 255, "alpha_transparent_fraction": 0.0, "alpha_opaque_fraction": 0.9984118938446045, "alpha_soft_fraction": 0.0015881061553955078, "alpha_thresholds": [5, 250], "rgb_dx_mean": 0.018379991874098778, "rgb_dy_mean": 0.018215516582131386, "decode_latents": "bird_background-s1000042-decode.pt", "decode_latents_sha256": "19746d2cebd5ebb888e1df24d7c2c07631fbb91fcf6f6650225caa6e84cb2246"}

evaluation/editing-v2/audit/probe/records.jsonl

@@ -0,0 +1,3 @@
+{"case": "cup_material", "size": 2048, "seed": 1000042, "steps": 40, "vae_tiling": true, "seconds": 205.40556079999988, "image": "cup_material-2048-s1000042.png", "sha256": "03c296f4cda7f737b068f6adb493265e1e207799ce3c92bd818999c34a70eee6"}
+{"case": "bird_background", "size": 2048, "seed": 1000042, "steps": 40, "vae_tiling": true, "seconds": 187.8806442, "image": "bird_background-2048-s1000042.png", "sha256": "177fc897274faa4846659ec273d8ae39453d67526dc91d2bcc71407ce4ccd9da"}
+{"case": "bird_background", "size": 1024, "seed": 1000042, "steps": 40, "vae_tiling": true, "seconds": 26.06802820000007, "image": "bird_background-1024-s1000042.png", "sha256": "7e51d43d20482ca2b42193b22bbe1285cd0a6175e5dc4f6338c813550b2d233f"}

evaluation/editing-v2/bf16/COMPLETE.json

@@ -0,0 +1 @@
+{"records": 10, "records_sha256": "41a5088800126b52312677a119585febdb0ed6a17191f1853667e8af19f8be28"}

evaluation/editing-v2/bf16/bird_background-s1000042-decode.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d60d06020654c11078db5eaec59f69f28eceefbee579b7959cf3d78494f6015
+size 526161

evaluation/editing-v2/bf16/bird_background-s1000123-decode.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e27fae36bf5bd455912505dbc8944a8e5bb01d5ff98641074891533a86d2a2d
+size 526161

evaluation/editing-v2/bf16/bird_remove_background-s1000042-decode.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:322370af42fb748668fb34d55f71b625b3b36f0ecb968836a34e702389c1ad0f
+size 526210