SkyReels V3
Open multimodal video model — multi-reference subjects, audio-driven talking avatars, video-to-video.
About
SkyReels V3 is Kunlun Tech's Skywork AI team flagship video model, open-sourced January 29, 2026 with inference code and weights (github.com/SkyworkAI/SkyReels-V3). Built on a unified multimodal in-context learning framework, it natively handles three tasks: multi-subject video generation from 1–4 reference images (characters, objects, backgrounds fused coherently), audio-conditioned talking avatars (minute-level, multi-character, lip-synced), and video-to-video generation — plus video extension and start/end frame control inherited from the V2 diffusion-forcing line.
The honest picture: as a January 2026 release the ecosystem (tutorials, LoRAs, integrations) is still maturing compared to Wan or HunyuanVideo. Multi-reference consistency is its headline claim but remains hard to evaluate independently. Still, fully open weights for reference-faithful video plus talking avatars in one model is a combination no other open release offers.
What it can do
- Multi-reference (1–4 image) subject-consistent video
- Audio-driven talking avatars, multi-character
- Video-to-video generation
- Video extension and start/end frame control
- Open weights with technical report
Setup
Clone github.com/SkyworkAI/SkyReels-V3, install per README, download weights from Hugging Face or ModelScope; a limited-time free API was offered via apifree.ai.
Prerequisites
CUDA GPU; see repo for VRAM requirements and low-VRAM flags.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 54; Trending score 76/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.