Add release card and licenses

LICENSE-AC3D.txt

@@ -0,0 +1,40 @@
+Copyright (c) 2025 Snap Inc. All rights reserved.
+
+This sample code is made available by Snap Inc. for non-commercial, research purposes only.
+
+Non-commercial means not primarily intended for or directed towards commercial advantage or monetary compensation. Research purposes mean solely for study, instruction, or non-commercial research, testing or validation.
+
+No commercial license, whether implied or otherwise, is granted in or to this code, unless you have entered into a separate agreement with Snap Inc. for such rights.
+
+This sample code is provided as-is, without warranty of any kind, express or implied, including any warranties of merchantability, title, fitness for a particular purpose, non-infringement, or that the code is free of defects, errors or viruses. In no event will Snap Inc. be liable for any damages or losses of any kind arising from this sample code or your use thereof.
+
+Any redistribution of this sample code, including in binary form, must retain or reproduce this license text including all copyright notices, conditions and disclaimers.
+
+The following sets forth attribution notices for third-party software that may be included in portions of this sample code:
+
+CogVideoX-ControlNet
+https://github.com/TheDenk/cogvideox-controlnet
+
+Copyright © 2024 CogVideo Model Team @ Zhipu AI
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
+
+CogVideoX
+https://github.com/THUDM/CogVideo/tree/main
+
+Copyright © 2024 CogVideo Model Team @ Zhipu AI
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
+
+CameraCtrl
+https://github.com/hehao13/CameraCtrl
+
+Copyright © 2025 Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, Ceyuan Yang
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

LICENSE.md

@@ -0,0 +1,21 @@
+# License and attribution
+
+This repository does not have one blanket license. Each checkpoint is governed by the terms associated with its implementation and base model.
+
+| Checkpoint family | Applicable terms |
+| --- | --- |
+| `gimbal_ac3d` | Snap AC3D non-commercial research license reproduced in `LICENSE-AC3D.txt`; CogVideoX and CogVideoX-ControlNet Apache-2.0 notices are included there. |
+| `gimbal_ucpe*` | UCPE code is MIT; the required Wan 2.1 base model and its weights retain the terms published by Wan-AI. |
+| `gimbal_wan_2_1` | DiffSynth-Studio and Wan 2.1 are published under Apache-2.0; the upstream terms continue to apply. |
+| `gimbal_wan_2_2*` | DiffSynth-Studio and Wan 2.2 are published under Apache-2.0; the upstream terms continue to apply. |
+
+GimbalDiffusion source code authored by the project is MIT licensed in the code repository. This notice does not replace or weaken any upstream restriction, and no rights are granted for base-model weights that are not distributed here.
+
+Upstream sources:
+
+- AC3D: https://github.com/snap-research/ac3d
+- UCPE: https://github.com/chengzhag/UCPE
+- DiffSynth-Studio: https://github.com/modelscope/DiffSynth-Studio
+- CogVideoX-2b: https://huggingface.co/THUDM/CogVideoX-2b
+- Wan 2.1: https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B
+- Wan 2.2: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B

README.md

@@ -0,0 +1,75 @@
+---
+license: other
+license_name: mixed-upstream-licenses
+license_link: https://huggingface.co/lefreud/GimbalDiffusion/blob/main/LICENSE.md
+pipeline_tag: text-to-video
+base_model:
+- THUDM/CogVideoX-2b
+- Wan-AI/Wan2.1-T2V-1.3B
+- Wan-AI/Wan2.2-TI2V-5B
+datasets:
+- YOUSIKI/PanoWan
+tags:
+- video-generation
+- camera-control
+- camera-trajectory
+- gravity
+- gimbaldiffusion
+---
+
+# GimbalDiffusion
+
+Paper-release checkpoints for **GimbalDiffusion: Gravity-Aware Camera Control for Video Generation** (ECCV 2026).
+
+- [Project page](https://lvsn.github.io/GimbalDiffusion/)
+- [Paper](https://arxiv.org/abs/2512.09112)
+- [Code](https://github.com/lefreud/GimbalDiffusion-dev)
+
+These are weights-only checkpoints for the custom backends in the code repository. They are not standalone Diffusers pipelines; download the matching base model and use the corresponding YAML configuration.
+
+## Checkpoints
+
+| Path | Backbone | Variant | Training steps | Bytes | SHA-256 |
+| --- | --- | --- | ---: | ---: | --- |
+| `gimbal_ac3d/controlnet.pt` | CogVideoX-2b / AC3D | Main | 10,000 | 21,839,151 | `b136e9afd08f41e0497b89fd0d3e34c2e41f7604cf4c9d081462ff783f869099` |
+| `gimbal_ucpe/dit.safetensors` | Wan 2.1 / UCPE | Main | 24,000 | 2,909,385,032 | `8c04d9e439d58259b0fad6cfeeed0cb5f0ba611ba278b24a9c234805133f7c4c` |
+| `gimbal_ucpe_no_np/dit.safetensors` | Wan 2.1 / UCPE | Without null-pitch conditioning | 24,000 | 2,909,385,032 | `aeabf392a7b30c62f283d41280a53ffc2580997d284d371eabac476a0269f3fb` |
+| `gimbal_ucpe_no_absolute/dit.safetensors` | Wan 2.1 / UCPE | Relative poses | 24,000 | 2,909,385,032 | `e88bfaf60e1916b3f6d09d7264880dff006009522eb4bb9296658a71975467dd` |
+| `gimbal_wan_2_1/dit.safetensors` | Wan 2.1 | Main | 30,000 | 2,941,896,272 | `e1259983d400e61ac4209953684fc0633030e2d129bfc2203cbc6fbe706c09aa` |
+| `gimbal_wan_2_2/dit.safetensors` | Wan 2.2 | Main | 30,000 | 10,006,745,128 | `427cb8d5d068d2e2811f45aa189f5e90e214b4a895c4bc87db3f2bc328ad10a0` |
+| `gimbal_wan_2_2_no_np/dit.safetensors` | Wan 2.2 | Without null-pitch conditioning | 30,000 | 10,006,745,128 | `4effd00f013f9a08fc6c8330bf4f2a7616c74771a51a7e58e903a7f861ac21a5` |
+
+## Download and use
+
+```bash
+hf download lefreud/GimbalDiffusion --local-dir checkpoint/gimbaldiffusion
+hf download THUDM/CogVideoX-2b --local-dir checkpoint/backbones/THUDM/CogVideoX-2b
+hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir checkpoint/backbones/Wan-AI/Wan2.1-T2V-1.3B
+hf download Wan-AI/Wan2.2-TI2V-5B --local-dir checkpoint/backbones/Wan-AI/Wan2.2-TI2V-5B
+python -m src.scripts.inference_gimbal --model gimbal_wan_2_2 --preset backflip
+```
+
+The code repository documents separate AC3D, UCPE, and WAN environments because UCPE and DiffSynth-Studio install incompatible packages named `diffsynth`.
+
+## Training data and evaluation
+
+The models were trained from PanoWAN clips with VIPE camera poses and gravity estimates. The release includes separate repositories for the [PanoWAN annotations](https://huggingface.co/datasets/lefreud/GimbalDiffusion-PanoWAN-metadata), [SpatialVID-Extreme](https://huggingface.co/datasets/lefreud/GimbalDiffusion-SpatialVID-Extreme), and [prompt-entanglement benchmark](https://huggingface.co/datasets/lefreud/GimbalDiffusion-Prompt-Entanglement). See the paper and supplement for the full evaluation protocol and results.
+
+## Limitations
+
+The checkpoints inherit the capabilities and limitations of their base video models. Camera control can fail for unusual trajectories, long durations, ambiguous prompts, or scenes with weak geometric cues. Generated videos may reproduce biases or unsafe content learned by the base models. These research checkpoints are not validated for production or safety-critical use.
+
+## License
+
+This repository contains artifacts governed by different upstream terms. In particular, the AC3D checkpoint is restricted to non-commercial research use. Read [LICENSE.md](LICENSE.md) and [LICENSE-AC3D.txt](LICENSE-AC3D.txt) before use. Base-model weights are not redistributed here and retain their own licenses.
+
+## Citation
+
+```bibtex
+@inproceedings{fortierchouinard2026gimbaldiffusion,
+  title = {GimbalDiffusion: Gravity-Aware Camera Control for Video Generation},
+  author = {Fortier-Chouinard, Fr{\'e}d{\'e}ric and Hold-Geoffroy, Yannick and Deschaintre, Valentin and Gadelha, Matheus and Lalonde, Jean-Fran{\c{c}}ois},
+  booktitle = {European Conference on Computer Vision (ECCV)},
+  year = {2026}
+}
+```