Storytellers and Filmmakers
Plan continuous shots with a clear opening, evolving action, and a stable setting before developing a finished sequence.
Tencent MovieRoll can build continuous video scenes from text or an image, then introduce new actions through ordered event prompts. Explore its long-form approach to keeping subjects, movement, and settings connected.
Explore six official MovieRoll examples, then edit a starting prompt to develop your own scene. The first three show Base; the next three show Distilled.
Tencent MovieRoll
A clownfish swimming in midair
Tencent MovieRoll
A panda playing guitar
Tencent MovieRoll
A giraffe on lifeguard duty
Tencent MovieRoll
Fishing for stars
Tencent MovieRoll
Sculpting flowing water
Tencent MovieRoll
A woman climbing an indoor wall
A video generation framework designed to carry a scene forward over time.
Tencent Hunyuan and Peking University released MovieRoll code on November 16, 2026, with contributors from HKUST. The project develops a Base model and a few-step Distilled model from HunyuanVideo 1.5.
Instead of treating a longer scene as unrelated short clips, MovieRoll rolls generation through successive windows. Its official showcases include 30-second scenes, while reported evaluations cover settings up to 60 seconds. These examples describe the research model, rather than a universal duration limit.
Tencent MovieRoll brings together three ways to direct a scene:
MovieRoll-Base supports 480p with arbitrary aspect ratios. The released Distilled checkpoints list 480 × 864 and 864 × 480 outputs.
MovieRoll uses the Tencent Hunyuan Community License. Review its terms and the rights to your inputs before using outputs.
Six parts of MovieRoll help connect the opening scene to the actions that follow.
MovieRoll generates successive windows of video, with training and inference designed to reduce degradation during longer rollouts.
Build a scene around written descriptions of subjects, camera placement, lighting, and sustained motion.
Released image-to-video checkpoints let a visual starting point guide the scene before its movement develops.
Ordered prompts introduce new actions while carrying the setting forward, giving a sequence more than one narrative beat.
MovieRoll-Base uses 50 denoising steps; the released Distilled checkpoints use five. Step counts describe the models, rather than guaranteed generation times.
Hierarchical caching and cache recomputation retain useful scene history as the video progresses through successive windows.
Explore sustained action and connected scenes for creative work, communication, and research.
Plan continuous shots with a clear opening, evolving action, and a stable setting before developing a finished sequence.
Explore playful character scenes, recurring motion, and imaginative environments for visual storytelling.
Develop scene concepts from a written brief or visual starting point, then compare different actions and compositions.
Study single-event and multi-event generation, comparing Base and Distilled behavior over longer rollouts.
Turn a scene idea into an editable brief, then refine it in the video generator.
Use an example above or write a subject, setting, and action. Keep the camera direction clear and the motion easy to sustain.
Review the selected model in the generator and choose its aspect ratio, resolution, and duration. Keep the settings suited to the scene you want to make.
Submit when your brief is ready, review the result, and adjust the action or framing. Using an example only fills the prompt; it does not start generation.
Verified specifications of the released MovieRoll research checkpoints and official evaluation settings.
| Developer | Tencent Hunyuan and Peking University; HKUST contributors |
| Code release | November 16, 2026 |
| Model lineage | HunyuanVideo 1.5 |
| Inputs | Text-to-video and image-to-video checkpoints |
| Event control | Single prompt or ordered event prompts |
| Base output | 480p; arbitrary aspect ratios |
| Distilled output | 480 × 864 or 864 × 480 |
| Denoising steps | Base: 50; Distilled: 5 |
| Reported evaluation durations | Single event: 10, 30, 60 seconds; multi-event: 30, 60 seconds |
| Output format | Native inference supports MP4 video |
| License | Tencent Hunyuan Community License |
Last updated: November 17, 2026
Compare the two released editions of Tencent MovieRoll using the official checkpoint specifications.
| Specification | MovieRoll-Base | MovieRoll-Distilled |
|---|---|---|
| Input tasks | Text-to-video and image-to-video | Separate text-to-video and image-to-video checkpoints |
| Output dimensions | 480p with arbitrary aspect ratios | 480 × 864 or 864 × 480 |
| Denoising steps | 50 | 5 |
| Guidance | AnnealCFG | No classifier-free guidance |
| Reported evaluation durations | 10 and 30 seconds | 10, 30, and 60 seconds |
MovieRoll-Base Advantages: More flexible aspect ratios at 480p, with the full 50-step denoising process.
MovieRoll-Distilled Advantages: A five-step generation process and official evaluation results at 60 seconds.
Explore Base for flexible framing and Distilled for few-step rollouts. Compare output quality on your own scene; denoising steps alone do not determine visual quality or elapsed time.
Choose a Picir AI plan for video creation. The generator shows the selected model and its credit cost for your current settings.
Answers about MovieRoll scenes, inputs, specifications, examples, and creative workflows.
Explore the Tencent MovieRoll AI Video Generator page on Picir AI. Browse examples and use a starting prompt in the video generator.
MovieRoll is a framework from Tencent Hunyuan and Peking University for rolling out long-form videos. It includes Base and Distilled editions and builds on HunyuanVideo 1.5.
Yes. The released model family includes image-to-video checkpoints as well as text-to-video checkpoints. An image establishes the visual starting point for the generated scene.
The official project evaluates single-event clips at 10, 30, and 60 seconds and multi-event clips at 30 and 60 seconds. These are evaluation settings, rather than a published maximum duration. Review the selected generator model for the duration available in your current workflow.
MovieRoll-Base lists 480p with arbitrary aspect ratios. The released Distilled checkpoints list 480 × 864 and 864 × 480. Generator resolution choices follow the model currently selected in the tool.
An ordered sequence of event prompts changes the action as the video continues. Keep subjects and setting consistent, and make each event a clear continuation of the previous one.
No. They fill an editable starting prompt and supported settings. The official videos demonstrate MovieRoll; a starting prompt is a creative brief, rather than a promise to reproduce the same frames.
Plans and credits follow the options shown on Picir AI. Review the selected model and estimated credit cost in the generator before submitting a request.
Describe one setting, a clear subject, and motion that can continue. Specify a fixed camera or a deliberate tracking shot. For several events, preserve the scene while adding a new action to each prompt.
Use inputs you own or have permission to use, and follow content rules. Review the result and any applicable model license terms before publication or commercial use.