Style-Invariant Error Recycling
Add a perceptual style consistency loss and train on diverse video datasets to prevent error-recycling mechanisms from misinterpreting artistic styles as artifacts.
Source paper
Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, Alexandre Alahi
Generated by AI. The paper’s authors did not write these.
Add a perceptual style consistency loss and train on diverse video datasets to prevent error-recycling mechanisms from misinterpreting artistic styles as artifacts.
Introduce a global identity anchor via a separate cross-attention block to allow diffusion transformers to maintain long-term subject continuity beyond short-term frame windows.
Refactor parallel video generation diffusion models into a causal streaming architecture using causal temporal attention and KV-caching for real-time output.
This paper introduces Stable Video Infinity (SVI), a framework designed to generate infinite-length videos with high consistency by addressing the problem of error accumulation in autoregressive models. The authors identify a 'hypothesis gap' where models trained on clean data fail when exposed to their own self-generated errors during inference, and they propose 'Error-Recycling Fine-Tuning' to train the model to actively correct these errors.
The core technique is Error-Recycling Fine-Tuning (ERFT). Instead of training on clean ground-truth transitions, the method injects historical errors (predicted from a replay memory bank) into the input latents during training. It uses a bidirectional integration step to calculate specific error terms (video latent error, noise error, and image condition error) and stores them in a dynamic memory bank. This forces the Diffusion Transformer (DiT) to learn a velocity field that points back to the clean trajectory from an error-corrupted state, effectively bridging the gap between training and autoregressive inference.