KeyID introduces training-free method for identity-preserving video generation

The paper KeyID has been published on arXiv — a training-free framework for identity-preserving video generation that separates video dynamics synthesis from identity injection. The authors propose first generating a draft video without tying it to a specific face, then correcting sparse key frames and interpolating motion between them. The method took second place in Track 2 (Sequential Action) at the ACM Multimedia 2026 IPVG Grand Challenge. arXiv

What it means

KeyID is a research framework, not a ready-to-use video generator you can work with directly. In our reference on AI for video, there are no tools that tackle identity-preserving video generation as their primary task. The closest tools by type are open-source video generators:

• LTX-Video — an image-to-video generator from Lightricks with open source. It works with images as input, but the fact sheet has no information about preserving character identity in long sequences of actions.
• CogVideoX — a text-to-video generator from Zhipu AI. It also supports image-to-video, but the task of keeping the same face in complex scenes is not covered in the fact sheet.
• Pyramid Flow — an open-source text-to-video generator. The description makes no mention of working with references or preserving identity.

KeyID is interesting for its approach: instead of dense control on every frame, the authors correct only key frames and interpolate motion between them. This reduces the conflict between following the text prompt and the accuracy of face reproduction. The developer does not specify VRAM requirements or the platforms on which the method was run. For now, KeyID remains a research project: the code is publicly available, but integration with popular libraries such as diffusers is not stated in the paper. Practical use in combination with existing video generators will require your own adaptation.