Surprise Forcing: New AI method boosts long video quality while cutting compute waste
A training-free fix remembers crucial frames and skips easy denoising steps automatically.
Long video generation with autoregressive diffusion models suffers from two resource-allocation problems. The rolling key-value cache forgets distant visual evidence, and every generated chunk receives the same number of denoising passes regardless of its actual difficulty. Surprise Forcing treats both as online allocation problems. A Surprise-Gated Memory Bank summarizes evicted frames with value-token descriptors, evaluates them using global-deviation and nearest-neighbor novelty signals, and regulates admission through a feedback-controlled budget. Priority-based replacement and relevance-aware routing keep the external memory compact and useful.
In parallel, Surprise-Aware Denoising estimates chunk difficulty from the maximum adjacent-frame cosine distance after the first denoising pass, then uses a local percentile scheduler to skip intermediate steps for comparatively easy chunks. Experiments on VBench, VBench-Long, and VBench-2.0 show that this allocation strategy improves long-horizon consistency and visual quality while retaining real-time streaming throughput. The method is training-free, meaning it can be applied to existing models without additional fine-tuning.
- Training-free framework compatible with any streaming autoregressive diffusion video model.
- Surprise-Gated Memory Bank uses global deviation and nearest-neighbor novelty to selectively retain important frames.
- Surprise-Aware Denoising skips intermediate denoising steps for easy chunks, improving efficiency without quality loss.
Why It Matters
Makes minute-scale, real-time video generation more practical by intelligently allocating compute to where it matters most.