ID-V2V restyles a video while preserving the original characters' identity and performance. Given a source video and a stylized keyframe (plus optional extra keyframes and a text prompt), it generates a new video whose scene, lighting, and style follow the keyframe(s), while the source subjects' identity, expressions, gaze, and motion are kept intact — a shoot first, restyle later workflow. It is a research artifact, released for demonstration and inspiration only.

This repository hosts the two finetuned ID-V2V checkpoints. Both share the same architecture (Wan 2.1 image-to-video with VACE control). See the code repository for how to use this model.

Checkpoints

File	Variant	Control conditions
`idv2v.pth`	Default (recommended)	1 — foreground-on-gray pixels (subject segmented by SAM3)
`idv2v_with_normal_depth.pth`	Alternate	3 — foreground-on-gray pixels + surface normals + depth

idv2v.pth — the default single-condition model. It preserves the segmented subject (relit to match the new scene) and regenerates the rest of the frame from the prompt.
idv2v_with_normal_depth.pth — additionally conditions on the source video's surface normals (DAViD) and depth (DepthAnything-V2), giving tighter geometric control over the regenerated regions.

The two are different, non-interchangeable weights with an identical architecture, so loading the wrong one does not raise an error — it silently degrades the output. Pair each checkpoint with its matching script, as documented in the code repository.