ID-V2V — Public project resources¶

Development line: project:id-v2v · thread public-project-resources
Last event: 2026-07-30 · 1 dated since 2026-07-30 · Researched: 2026-09-05 · confidence: high

What it is¶

ID-V2V is an open-source research pipeline built on Wan2.1 I2V-14B-720P and VACE for video editors. It transforms source video, a stylized first frame, optional keyframes, and a text prompt into a new video while preserving human appearance, expression, gaze, and motion.

• Restyling transfers new visual appearance to the footage.
• Relighting adjusts scene illumination while preserving identity.
• Multi-character handling processes several people in the same scene.
• Extra keyframes guide intermediate character generation.
• Long-video processing divides longer sequences across clips.

Generation runs in 720p, base weights and dependencies take about 96 GB, and the author testbed is 8× A100 80GB. This is a heavy research artifact for local reproducible inference, not a production service.

Development line¶

• 2026-07-30 — ID-V2V public project resources were linked. Official code defines the input directory, preprocess-to-generate workflow, and 720p generation. It processes clips longer than 81 frames using overlapping windows. It provides recipes for restyling, relighting, and keyframes. A model card update from 2026-07-29 notes the base architecture as Wan 2.1 image-to-video with VACE control. The weights are intended only for demonstration and inspiration.

What changed¶

2026-07-24 — The arXiv paper "ID-V2V: Identity-Preserving Video Restylization" was published, describing the separation of keyframe-guided synthesis and source-grounded identity preservation through relighting. 2026-07-29 — Hugging Face published two fine-tuned, incompatible checkpoints: the recommended idv2v.pth with one SAM3 foreground-on-gray condition and idv2v_with_normal_depth.pth with additional normal and depth controls. 2026-07-30 — The release became practically reproducible: official code defines an input directory, preprocess-to-generate workflow, 720p generation, processing clips longer than 81 frames via overlapping clips, and recipes for restyling, relighting, and keyframes. A model card update from 2026-07-29 notes the base architecture as Wan 2.1 image-to-video with VACE control, with weights meant only for demonstration and inspiration.

How to use this¶

As of 2026-07-30, practitioners should begin ID-V2V evaluation or reproduction from the linked official project page, source repository, and Hugging Face page, verifying exact capabilities and release status there before use.

1. Clone the repository, create a unified environment with uv sync, activate it, and log in to Hugging Face early because SAM3 requires gated access. — https://github.com/Eyeline-Labs/ID-V2V
2. Download standard checkpoints and dependencies with bash scripts/download_checkpoints.sh; add --with-depth for the normal and depth variant. — https://github.com/Eyeline-Labs/ID-V2V
3. Prepare a directory with source.mp4, stylized_first_frame.png, and prompt.txt; add keyframes/<frame_number>.png when needed. — https://github.com/Eyeline-Labs/ID-V2V
4. For the standard model, run preprocessing with the SAM3 text prompt, then run inference; relighting uses a separate path without preprocessing. — https://github.com/Eyeline-Labs/ID-V2V
5. For long source footage, add keyframes and use the longer-video recipe: the system joins clips across overlapping boundary frames. — https://github.com/Eyeline-Labs/ID-V2V

Best practices¶

• Match each checkpoint to its variant script: matching architecture raises no error on an invalid pair, but output silently degrades. — https://huggingface.co/Eyeline-Labs/ID-V2V
• Use more keyframes on long videos so drift between clips stays low. — https://github.com/Eyeline-Labs/ID-V2V
• Treat results as experimental, verifying identity preservation, facial expressions, and lip-sync on your own footage before production use. — https://huggingface.co/Eyeline-Labs/ID-V2V

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The provided schema omits event_findings and new_events; their verifiable details sit in what_changed.
• Official materials describe reproducible inference, but lack confirmed production support, a public hosted API, or minimum GPU requirements.

Sources¶

Agent brief¶

• Subject: project:id-v2v, thread public-project-resources, 1 dated events 2026-07-30 → 2026-07-30.
• Practical note: As of 2026-07-30, practitioners should begin ID-V2V evaluation or reproduction from the linked official project page, source repository, and Hugging Face page, verifying exact capabilities and release status there before use.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.