Style-Aware Reference Adapter for Non-Photorealistic domains
Integrate a style-aware adapter into motion generation models to modulate facial expression blocks and preserve the unique stylistic identity of non-humanoid characters.
Source paper
Shuyuan Tu, Yueming Pan, Yinming Huang, Xintong Han, Zhen Xing, Qi Dai, Kai Qiu, Chong Luo, Zuxuan Wu
Generated by AI. The paper’s authors did not write these.
Integrate a style-aware adapter into motion generation models to modulate facial expression blocks and preserve the unique stylistic identity of non-humanoid characters.
Replace heuristic scaling functions in latent acceleration mechanisms with a lightweight temporal neural network that directly regresses Taylor expansion residual errors.
Distill multi-step diffusion transformer models into low-step student architectures using sliding window consistency losses to achieve real-time video generation.
The paper presents FlashPortrait, a video diffusion transformer framework for generating infinite-length, identity-preserving portrait animations with improved inference speed. It introduces an Adaptive Latent Prediction mechanism that uses Taylor expansion to skip denoising steps and a Normalized Facial Expression Block to align facial features with diffusion latents, achieving up to 6x acceleration compared to baselines.
The model is based on the Wan2.1 Video Diffusion Transformer (DiT). It employs a Normalized Facial Expression Block to align the distribution statistics (mean/variance) of extracted facial features with diffusion latents. For acceleration, it utilizes a training-free Adaptive Latent Prediction mechanism that applies Taylor series expansion on historical latents within a sliding window to predict future latents, modulated by dynamic functions based on latent variation rates.