FreqEdit is a training-free framework for robust multi-turn image editing, preserving high-frequency features to maintain identity consistency, visual quality, and instruction-following across iterative edits.

An ODE-native GRPO framework that aligns autoregressive video generators through KV-based semantic exploration.

AnyTalker is an audio-driven multi-person talking video generation framework that supports scalable identity control, natural lip synchronization, and interactive conversational behaviors. It introduces Audio-Face Cross Attention and a two-stage training strategy to learn multi-person speaking patterns with low data cost.