[SIGGRAPH Asia 2026] 4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Third-party notices
The root Apache-2.0 license applies only to first-party 4DAnyone code. The following components and assets retain their own terms.
GVHMR
- Upstream source: https://github.com/zju3dv/GVHMR
- Pinned upstream revision:
6ec3ca39336c50492c0fae65fba2fb831fc7d866 - Location:
third_party/GVHMR - License:
third_party/GVHMR/LICENSE
4DAnyone uses the official upstream source without a GVHMR source patch. The pinned revision is immutable and can be fetched anonymously from the upstream repository.
GVHMR, HMR2, ViTPose, YOLO, SMPL, and SMPL-X model files are not bundled in 4DAnyone. The complete public inference closure is anchored by the immutable Hugging Face revision frozen in fdanyone/assets.py. Users obtain those files through the applicable upstream or licensed setup flows.
GVHMR permits use, copying, modification, and distribution for educational, research, and non-profit purposes only. It requires modifications based on the work to be open source and prohibits commercial use. Its checkpoints and transitive model assets may have additional terms.
BiRefNet
- Upstream source and model: https://huggingface.co/ZhengPeng7/BiRefNet
- Pinned model revision:
e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4 - Upstream code: https://github.com/ZhengPeng7/BiRefNet
- License: MIT, copied at
third_party/licenses/BIREFNET_LICENSE
The adaptive preprocessing path downloads the pinned BiRefNet configuration, inference source, and weights on demand. It uses the resulting foreground masks for source cropping and source-framing analysis. Those downloaded files retain their upstream terms.
PyTorch3D compatibility subset
- Upstream source: https://github.com/facebookresearch/pytorch3d
- Upstream revision:
f34104cf6ebefacd7b7e07955ee7aaa823e616ac - Location:
fdanyone/vendor/pytorch3d_compat - License: BSD 3-Clause, copied at
fdanyone/vendor/pytorch3d_compat/LICENSE
Classic GVHMR imports PyTorch3D rotation conversions through a broader training-time module surface. 4DAnyone retains the two upstream source files needed by inference, plus small first-party module aliases and a native PyTorch fallback for an imported optional KNN helper. The renderer and compiled PyTorch3D operators are not included or required. UPSTREAM.md and VENDORED_FILES.txt record the exact retained closure and source hashes.
Pexels example media
- Source platform: https://www.pexels.com/
- Location:
data/source/pexels - Terms: https://www.pexels.com/license/
The companion Hugging Face repository distributes 20 modified 121-frame excerpts for CLI inference and 4DAnyone Space. They are installed under the location above on demand. Their filenames, source videos, extraction ranges, and checksums are recorded in the pinned Pexels asset manifest. Pexels permits its media to be used and modified for free and does not require attribution. The source links are retained for provenance and creator credit.
DiffSynth-Studio inference runtime
- Public source: https://github.com/modelscope/DiffSynth-Studio
- Public base revision:
04e39f7de53df7276a7b40ca1791c2a393e05ff3 - Original experiment fork revision:
c00782d90c872c97bda4745a9e6a41a0a4a7c4db - Location:
fdanyone/vendor/diffsynth - License: Apache-2.0, copied at
fdanyone/vendor/diffsynth/LICENSE
UPSTREAM.md, UPSTREAM.patch, and VENDORED_FILES.txt in that directory record the exact provenance, research patch, and retained file set.
One retained DiffSynth file carries additional code-level upstream attribution:
models/wan_video_pose_encoder.pyderives its pose network from Tencent/MimicMotion, Copyright Tencent, under Apache-2.0.
Sapiens2-derived Goliath/MHR schema material
- Source project: https://github.com/facebookresearch/sapiens2
- Source revision:
0e51c12d7c7257d88431b2d50e523a7b03004854 - Source file:
sapiens/pose/configs/_base_/keypoints308.py - Retained material: the exact MHR70 name/order and Goliath40 link closure needed to reproduce model conditioning
- First-party Apache-2.0 material: RGB palette, color assignment policy, and renderer implementation in
fdanyone/skeleton - License copy:
third_party/licenses/SAPIENS2_LICENSE.md
No Sapiens2 model weights are distributed by this repository.
The retained schema material is conservatively treated as Sapiens2-derived and the complete agreement is included rather than assuming that the root Apache-2.0 license applies. The agreement contains explicit prohibited-use terms, including surveillance, biometric processing, identification or re-identification, deepfakes or deceptive content, and the other activities listed in section 1(b)(vi).
Model and body assets
The 4DAnyone checkpoint and sanitized MHR70 sparse regression tensors are first-party release artifacts licensed under Apache-2.0, published at the Hugging Face revision frozen in fdanyone/assets.py. The regressor file also embeds the exact keypoint-name order required to interpret those tensors. That schema material retains the Sapiens2 notice above and is not relicensed by the tensor license. The Wan VAE and UMT5-derived frozen prompt conditioning, GVHMR checkpoints, HMR2, ViTPose, detector weights, and SMPL-X body models remain third-party assets. The redistributable files are mirrored by the companion Hugging Face repository. Users obtain SMPL-X from its official site and remain responsible for every applicable license.
The frozen base-model assets come from the official Wan2.2-TI2V-5B repository at revision 37685f96025fc1425edccdd4b2bca3836ae917ff and the official Wan2.1-T2V-1.3B repository at revision 3f40b6dc4ca5c02dd23c9db74d9d2ccb82903b86. The latter is used offline to produce the fixed prompt-conditioning tensor; its T5 weights and tokenizer are not part of reader downloads. Their official model cards license those model assets under Apache-2.0. The redistributed copies are pinned by the frozen Hugging Face revision.
Wan2.2 TI2V 5B Turbo LoRA
4danyone/Wan22_TI2V_5B_Turbo_lora_rank_64_fp16.safetensors is a byte-for-byte mirror of the Rank64 difference-LoRA published by Kijai/WanVideo_comfy at revision 86c2b0442e01eeee630b48fd7efc0cd37af03252. Kijai identifies its source model as quanhaol/Wan2.2-TI2V-5B-Turbo, whose weights are published under CC BY-NC-SA 4.0. The adapter is redistributed under those upstream terms. Kijai declares no separate adapter license, and the safetensors metadata is empty. Its exact SHA-256 remains pinned in fdanyone/assets.py; the model card retains the full upstream file identity and provenance.
The MHR70 asset is a sanitized inference artifact containing only the sparse support/weight tensors, keypoint names, and a path-free structural schema. The first-party regression tensors are redistributed under Apache-2.0. The embedded name/order metadata remains covered by the schema notice above.
That regressor was trained by the 4DAnyone project on DNA-Rendering frames, using outputs from SAM 3D Body and the Momentum Human Rig conversion tools. SAM 3D Body is governed by the SAM License, MHR source is Apache-2.0, and DNA-Rendering access requires a separate signed dataset agreement. The regressor is not bundled in the source repository. Its Hugging Face model card must retain these source-model and dataset citations alongside the split tensor/schema license declaration in this notice.
VGG-19 perceptual reconstruction
- Loss reference: Crowdsampling the Plenoptic Function, released under MIT in its upstream README
- VGG-19 model: Oxford Visual Geometry Group
- MatConvNet distribution: imagenet-vgg-verydeep-19, source MD5
106118b7cf60435e6d8e04f6a6dc3657 - Model license: Creative Commons Attribution 4.0 International
The optional splatfacto-perceptual method uses a custom pixel-plus-feature objective rather than standard LPIPS. The companion Hugging Face repository contains a modified copy of the VGG-19 model: MatConvNet filters are transposed into PyTorch layout and serialized as safetensors, while the unused classifier and final two convolution layers are omitted. The retained tensor values are otherwise unchanged. Credit remains with Karen Simonyan and Andrew Zisserman. Users must preserve that attribution and the CC BY 4.0 license when redistributing the converted weights.