Netflix releases ID-V2V video restyling under Apache 2.0

• Netflix's ID-V2V restyles recorded scenes while retaining a performer's identity, expressions, gaze, motion, and lip sync, according to the launch demo.
• A styled first frame can miss the performer's pose, yet later frames return to the source performance in the keyframe example.
• Relighting is a dedicated use case, and minchoi says in the relighting example that it skips normal SAM3 preprocessing.
• Code and model weights are public under Apache 2.0, and the release post says 720p output can extend through stitched overlapping clips.

Eyeline's project page begins with a source video and a stylized keyframe. The GitHub repository calls the release a research exploration, while its Hugging Face model card describes two checkpoint variants.

ID-V2V's official model card assigns the inputs separate jobs:

• Source video: the original performer and their motion.
• Stylized first frame: the target scene layout, lighting, and visual treatment.
• Optional extra keyframes and text prompt: additional direction for the generated sequence.

The model card calls this “shoot first, restyle later,” the same shorthand used in minchoi's workflow post.

The official repository divides the process into preprocessing, which derives control signals from the source clip, and generation, which combines those controls with the styled-frame direction. The output aims to change the world around the actor without reanimating the performance.

Longer output comes from overlapping 81-frame clip renders that are stitched into a single video, according to the GitHub repository. The release post identifies 720p as the output resolution.

The model card lists two finetuned checkpoints on a Wan 2.1 image-to-video architecture with VACE control:

• idv2v.pth: the default checkpoint uses a SAM-segmented foreground on gray, relights the performer, and regenerates the rest of the image from the prompt.
• idv2v_with_normal_depth.pth: an alternate checkpoint that adds source-video surface normals from DAViD and depth from DepthAnything-V2 for tighter geometric conditioning.

The checkpoints require their matching scripts. The card warns that mixing them silently degrades results.

bennash's Escher experiment uses GPT-6 Astra/Codex and Minimax H3, rather than ID-V2V, to turn an M.C. Escher print into a textured camera fly-through in under an hour. The thread breaks the workflow into six moves:

1. Built a 3D scene from the source image.
2. Added surrounding arena geometry for all-angle backgrounds.
3. Rendered a camera move around and through the building.
4. Split the video into three segments.
5. Generated a photoreal, 1980s-fantasy-film texture reference from the print.
6. Restyled each video segment against that texture in Minimax H3.

The published code is public, but its disclosed test scale is substantial: minchoi said in the hardware reply that the research team tested ID-V2V on eight A100 80GB GPUs.