Autoregressive Diffusion
Overview
Autoregressive diffusion combines sequential prediction of frames, chunks, tokens, or motion states with diffusion denoising, aiming to pair temporal control with detailed generation. Common approaches include latent diffusion transformers trained with teacher forcing, self-forcing, or distillation, alongside inference methods for memory selection, adaptive denoising, and speculative decoding. Work spans long-form video, language, motion, audio, and physical simulation, with evaluations including VBench, HumanML3D, and long-rollout stability tests. A central challenge is preventing errors from accumulating when models condition on their own imperfect outputs while keeping latency and memory manageable; causal generation also competes with the global coordination and speed of bidirectional methods.
Latest papers
52 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.