Latent-Space Saliency Distillation
Train a lightweight Perceptual Distillation Network operating directly on latent tokens to approximate VAE decoder saliency without expensive decoding steps.
Source paper
Zhibin Qin, Zhenxiong Tan, Zeqing Wang, Songhua Liu, Xinchao Wang
Generated by AI. The paper’s authors did not write these.
Train a lightweight Perceptual Distillation Network operating directly on latent tokens to approximate VAE decoder saliency without expensive decoding steps.
A method that calculates dynamic thresholds per image using statistical distributions to improve token classification efficiency in video generation.
A system that extends spatial token caching to 3D spatiotemporal tubes, skipping redundant computations in static background regions across video clips.
This paper introduces SpotEdit, a training-free framework designed to accelerate instruction-based image editing in Diffusion Transformers (DiTs). By dynamically identifying non-edited regions using a perceptual similarity metric and selectively reusing cached features for those areas, SpotEdit avoids redundant computation and preserves background fidelity.
The model uses 'SpotSelector', which calculates an LPIPS-like score from VAE decoder features to classify tokens as edited or non-edited during the diffusion process. It then employs 'SpotFusion', a mechanism that blends cached features from non-edited regions with condition image features using a time-dependent interpolation, allowing the model to skip DiT block computations for stable regions.