Video Models

Tencent Hunyuan Releases MovieRoll for 60-Second Video Generation

November 17, 20260 views阅读中文原文

Tencent Hunyuan and collaborators released MovieRoll's code and weights to address the loss of coherence and quality in long generated videos. The release includes a 50-step base model and 5-step distilled text-to-video and image-to-video variants. The team reports up to 60 seconds of continuous generation.

Hardware and output

The published setup generates 480 × 864 or 864 × 480 video at 24 fps. A complete single-GPU pipeline calls for 96 GB of VRAM, and the documented FlashAttention-3 setup targets Hopper-class hardware. Consumer GPU support was still on the roadmap in the release covered by the article.

Where it fits

MovieRoll is useful for studying long-video consistency and producing low-resolution previews. The release supports sequenced prompts for changing events in one shot. Benchmark tables in the original article distinguish durations, distilled and base checkpoints, and single- versus multi-event scenes. The long-duration results are from the project's own evaluations.

Check access and licensing

The weights are gated, and the project uses a Hunyuan community license. Confirm the current terms and regional availability before deploying. The source article recommends starting with a short distilled run to validate the environment and result quality before attempting a full 60-second clip.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...