StableVideo

Why we included this project

StableVideo is a research implementation for changing what appears in an existing video, such as repainting an object, recoloring a car, or altering a scene. It edits a layered representation of the footage once and propagates that change across frames, which keeps the result temporally consistent instead of flickering or drifting between shots. That makes it a useful reference for researchers reproducing the ICCV 2023 paper, and for developers curious whether prompt-driven video editing can fit their own pipelines. The repo includes a Gradio app, a CPU-only variant, and detailed VRAM tables, so you can estimate the memory cost before committing hardware. Expect a research-grade setup: you will need pretrained ControlNet weights and example video atlases, and the default float32 configuration wants about 29 GB of VRAM, though the amp and CPU-cache flags can cut that nearly in half.