ID-V2V: Identity-Preserving Video Model Gets ComfyUI Support via Kijai

What is ID-V2V?

ID-V2V (Identity-Preserving Video-to-Video Generation) is a research model accepted at SIGGRAPH Asia 2026, developed by Eyeline-Labs in collaboration with Netflix. It restyles a video's scene, lighting, and style while preserving the original characters' identity, expressions, gaze, and motion — a shoot first, restyle later workflow.

Given a source video and a stylized keyframe (plus optional extra keyframes and a text prompt), ID-V2V generates a new video whose background and aesthetics follow the keyframe, while the subjects' identity and performance remain intact.

Key details:

• Base model: Wan-AI/Wan2.1-I2V-14B-720P with VACE control
• License: Apache 2.0
• Paper: arXiv:2607.22830 (SIGGRAPH Asia 2026)
• Code: github.com/Eyeline-Labs/ID-V2V

Paul Debevec's Involvement

The project includes Paul Debevec — a pioneering computer graphics researcher who holds a Scientific and Technical Academy Award, a Lifetime Achievement Emmy, and a VES Fellowship. Debevec invented HDR imaging, image-based lighting, and the Light Stage facial capture system used in blockbuster films from Spider-Man to Avatar.

For ID-V2V, Debevec's expertise in relighting is critical: keeping a character's identity intact while correctly lighting them in a fully regenerated scene is the project's hardest challenge, and it builds directly on his prior work (Lux Post Facto at CVPR 2025, DifFRelight at SIGGRAPH Asia 2024).

ComfyUI Integration

On July 29, Kijai submitted PR #15139 to the ComfyUI core repository, adding the necessary infrastructure to support ID-V2V. The changes include:

1. VACE + I2V base support — the model uses Wan 2.1 I2V as its base with VACE control layers, a combination not previously supported natively in ComfyUI
2. Reference pad image input — a new input to the I2V node for the stylized keyframe that ID-V2V uses as the style target
3. Int8 convrot weights — available at Kijai/Wan_ID_V2V_comfy for VRAM-efficient inference

Kijai described the integration as relatively straightforward ("only +17 lines of code"), and the PR includes a ready-to-use workflow JSON. Community member T2 also independently built a wrapper node for the model.

Checkpoints

ID-V2V offers two checkpoints on Hugging Face:

Demo

The project page showcases several demos including background replacement, non-aligned keyframe stylization, and relighting:

Background replacement demo from the ID-V2V project page.

Note: ID-V2V is a research artifact released for demonstration and inspiration. It requires significant VRAM (~16GB+) and is best suited for users familiar with running Wan 2.1 I2V and VACE models in ComfyUI.