You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

MovieRoll is made available under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License"; see the LICENSE file in this repository). The License applies in the Territory only β€” worldwide excluding the European Union, the United Kingdom, and South Korea β€” and no rights are granted under the License in those territories. By clicking Agree, you represent and warrant that you are not located in the European Union, the United Kingdom, or South Korea, and that you will not access, use, redistribute, or make available MovieRoll (or any Model Derivative or Output thereof) in those territories. You also agree to be bound by the terms of the License, including its Acceptable Use Policy.

Log in or Sign Up to review the conditions and access this model content.

MovieRoll: And... Action!

A Systematic Framework for Rolling Out Long-Form Videos

Yunyang Ge1,2,*,‑, Weijie Kong2,*, Ryan Xu2,*, Yue Wu2, Qi Tian2, Jiangfeng Xiong2
Bin Lin1,2, Mingzhe Zheng2,3, Shenghai Yuan1, Zheyuan Liu1,2, Jianwei Zhang2,Β§, Li Yuan1,†
1 Peking University  Β·  2 Tencent Hunyuan  Β·  3 The Hong Kong University of Science and Technology
* Equal contribution  Β·  Β§ Project leader  Β·  † Corresponding author
‑ Work done during internship at Tencent Hunyuan

arXiv Project Page Models SGLang vLLM-Omni License

This repository hosts the MovieRoll transformer weights for Base and Distilled text-to-video / image-to-video models. Code, training recipes, and inference adapters are available on GitHub.

⚠️ License and territory notice. MovieRoll is released under the Tencent Hunyuan Community License Agreement (see LICENSE). The license is expressly limited to the Territory β€” worldwide excluding the European Union, the United Kingdom, and South Korea β€” and does not apply in those regions. Access is not granted to users located in the European Union, the United Kingdom, or South Korea.

Download Weights

Accept this repository's access conditions, then authenticate before downloading:

hf auth login

MovieRoll Checkpoint

Choose base, distilled/t2v, or distilled/i2v:

MODEL=distilled/t2v
hf download yunyangge/MovieRoll --include "$MODEL/*" --local-dir ./ckpts/movieroll

Each checkpoint folder contains config.json and model.safetensors.

Shared VAE and Text Encoders

All three MovieRoll checkpoints require the following shared weights:

hf download tencent/HunyuanVideo-1.5 --include 'vae/*' --local-dir ./ckpts
hf download Qwen/Qwen2.5-VL-7B-Instruct --local-dir ./ckpts/text_encoder/llm
hf download google/byt5-small --local-dir ./ckpts/text_encoder/byt5-small
modelscope download --model AI-ModelScope/Glyph-SDXL-v2 --local_dir ./ckpts/text_encoder/Glyph-SDXL-v2

Quick Start

First clone the MovieRoll code repository and follow its environment setup guide. Run the weight download commands above from the code repository root so that the checkpoints are placed under ./ckpts.

With distilled/t2v and the shared weights downloaded, run the following command from that same directory to generate a 30-second video on one GPU. The full pipeline must fit in GPU memory (a 96 GB GPU holds it). The native entry point does not yet support CPU offload.

python generate_movieroll.py \
  --model ckpts/movieroll/distilled/t2v --sp-size 1 \
  --prompt "In a softly lit, modern living room, a surreal event unfolds as a single orange and white clownfish swims gracefully through the air, continuously circling the head of a young man. The man, in his early twenties with dark, curly hair, sits calmly at a wooden desk, wearing a comfortable cream-colored knit sweater. His expression is one of quiet awe, his eyes subtly tracking the fish's rhythmic, looping path around him. The clownfish moves as if through water, its fins gently propelling it forward in a steady, circular motion. Its scales glisten with a wet sheen under the warm, diffused light coming from a large window to the left, and a few stray, glistening water droplets seem to trail in its wake, hanging in the air for a moment before vanishing. In the background, intentionally out of focus, sits a rectangular glass fish tank filled with clear blue water and swaying green aquatic plants, the source of this airborne marvel. The camera remains static in a medium shot, maintaining an eye-level perspective with the seated man. This fixed viewpoint captures both the man's serene reaction and the fish's complete, unending orbit around his head, emphasizing the dreamlike and continuous nature of the scene. Cinematic photography realistic style." \
  --seed 4890 --video-length 721 \
  --output-dir ./outputs

The video is saved to ./outputs/samples/default/rank0_sample_0.mp4. For image-to-video, multi-event prompts, and multi-GPU inference, see the inference options.

Citation

@misc{ge2026movierollsystematicframeworkrolling,
      title={MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos},
      author={Yunyang Ge and Weijie Kong and Ryan Xu and Yue Wu and Qi Tian and Jiangfeng Xiong and Bin Lin and Mingzhe Zheng and Shenghai Yuan and Zheyuan Liu and Jian-Wei Zhang and Li Yuan},
      year={2026},
      eprint={2611.21339},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.21339},
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for yunyangge/MovieRoll

Finetuned
(16)
this model

Paper for yunyangge/MovieRoll