Add perceptual VGG-19 convolution weights

LICENSE

@@ -34,9 +34,11 @@ Sapiens2-derived schema metadata stored in the same `.pt` file.
 | `gvhmr/epoch=10-step=25000.ckpt` | HMR2.0a from [4DHumans](https://github.com/shubham-goel/4D-Humans) | [MIT](licenses/4DHUMANS-MIT.txt) |
 | `gvhmr/vitpose-h-multi-coco.pth` | [ViTPose](https://github.com/ViTAE-Transformer/ViTPose) | [Apache-2.0](licenses/APACHE-2.0.txt) |
 | `gvhmr/yolov8x.pt` | [Ultralytics YOLOv8](https://github.com/ultralytics/ultralytics) | [AGPL-3.0](licenses/ULTRALYTICS-AGPL-3.0.txt) |
+| `perceptual/imagenet-vgg-verydeep-19-conv.safetensors` | [Oxford VGG-19](https://www.robots.ox.ac.uk/~vgg/research/very_deep/), converted from the [MatConvNet distribution](https://www.vlfeat.org/matconvnet/pretrained/) | [CC BY 4.0](licenses/VGG19-CC-BY-4.0.md) |
 | `data/source/pexels/*.mp4` | Pexels example clips; individual sources are listed in [`data/source/pexels/README.md`](data/source/pexels/README.md) | [Pexels License](https://www.pexels.com/license/) |
 
-The complete bundled license texts are in [`licenses/`](licenses/). Retain the
-applicable license, copyright, and attribution notices when redistributing an
-asset. Third-party assets remain governed by their upstream terms; their
-presence does not change the license of the first-party 4DAnyone checkpoint.
+The applicable license texts and notices are in [`licenses/`](licenses/).
+Retain the applicable license, copyright, and attribution notices when
+redistributing an asset. Third-party assets remain governed by their upstream
+terms; their presence does not change the license of the first-party 4DAnyone
+checkpoint.

README.md

@@ -38,12 +38,19 @@ gvhmr/
 ├── vitpose-h-multi-coco.pth
 └── yolov8x.pt
 
+perceptual/
+└── imagenet-vgg-verydeep-19-conv.safetensors
+
 data/source/pexels/
 └── eight 121-frame example videos
 ```
 
+The perceptual asset contains the first 14 convolution layers used by the
+LongVolCap reconstruction loss; the classifier and two unused final
+convolutions are omitted.
+
 Licensing is asset-specific. See [LICENSE](LICENSE) for the exact scope and
-[licenses/](licenses/) for the complete terms.
+[licenses/](licenses/) for the applicable terms and notices.
 
 ## Citation
 

licenses/VGG19-CC-BY-4.0.md

@@ -0,0 +1,30 @@
+# Oxford VGG-19 model notice
+
+`perceptual/imagenet-vgg-verydeep-19-conv.safetensors` derives from the
+19-layer model released by Karen Simonyan and Andrew Zisserman of the Visual
+Geometry Group, University of Oxford:
+
+- Project and attribution: <https://www.robots.ox.ac.uk/~vgg/research/very_deep/>
+- Paper: *Very Deep Convolutional Networks for Large-Scale Image Recognition*,
+  Karen Simonyan and Andrew Zisserman, ICLR 2015
+- MatConvNet source file:
+  <https://www.vlfeat.org/matconvnet/models/imagenet-vgg-verydeep-19.mat>
+- Source MD5: `106118b7cf60435e6d8e04f6a6dc3657`
+- License: [Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/legalcode)
+
+The Oxford project page states that the VGG-16 and VGG-19 models are released
+under the Creative Commons Attribution License and links specifically to CC BY
+4.0. CC BY 4.0 permits sharing and adaptation, including commercially, provided
+appropriate credit and a license link are retained and modifications are
+identified.
+
+## Modifications
+
+The source MatConvNet `.mat` file was converted into safetensors for direct,
+pickle-free PyTorch loading. Its convolution filters were transposed from
+MatConvNet HWIO layout to PyTorch OIHW layout. Only the first 14 convolution
+layers (through `conv5_2`) evaluated by the LongVolCap perceptual loss are
+retained; the classifier and final two unused convolution layers are omitted.
+The retained floating-point tensor values are otherwise unchanged.
+
+No endorsement by the original authors or the University of Oxford is implied.

perceptual/imagenet-vgg-verydeep-19-conv.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0beaf06caed0d0b752022c756bdf1752e941652ea30cf91e593438797f3293db
+size 61221640