SketchVideo: sketch-based video generation and editing

SketchVideo aims to achieve sketch-based spatial and motion control for video generation, and supports fine-grained editing of real or synthetic videos.

Model Details

Model Description

SketchVideo achieves sketch-based video clip generation (~6 seconds, 8 fps) based on text prompts, one or two keyframe sketches (at arbitrary time point).

SketchVideo also achieves sketch-based editing for video clips (~6 seconds, 8 fps) with the same inputs.

This model was trained to generate 49 video frames at a resolution of 720x480 given a context frame of the same resolution.

• Model type: Video Diffusion Model
• Finetuned from model: CogVideo-2b (720x480)

Model Sources

For research purpose, we recommend our Github repository (https://github.com/IGLICT/SketchVideo),
which includes detailed implementations.

• Repository: https://github.com/IGLICT/SketchVideo
• Paper: https://arxiv.org/abs/2503.23284
• Project page: http://geometrylearning.com/SketchVideo/
• Video: https://www.youtube.com/watch?v=eo5DNiaGgiQ

Uses

Feel free to use it under the Apache-2.0 license. Note that we don't have any official commercial product for ToonCrafter currently.

Limitations

• The generated videos are relatively short (6 seconds, FPS=8).
• For two keyframe sketches, they should have consistent content.

How to Get Started with the Model

Check out https://github.com/IGLICT/SketchVideo