New Steady-Forcing method keeps AI video backgrounds stable without killing motion

Autoregressive video diffusion models excel at generating short clips, but when asked to produce long, multi-minute sequences, they commonly suffer from two opposing failure modes: static scene layouts gradually drift away (spatial persistence failure), and attempts to lock down the background tend to suppress natural motion like water flow, fire flicker, or smoke dynamics (motion continuity failure). The team behind Steady-Forcing explicitly separates these problems by focusing on fixed-camera nature scenes, where background stability and fluid dynamics are more distinguishable than in moving-camera setups.

Their solution—Steady-Forcing—combines several innovations: a persistent visual anchor called V-Sink that remembers key background features, an exponential moving-average motion memory (EMA-Sink) that tracks natural motion patterns, block-relative temporal encoding to maintain temporal coherence, periodic cache purification to discard stale information, and distillation from a large Wan2.1-14B teacher model with motion-rewarded priors. In blind user studies, Steady-Forcing was preferred for both stability and motion continuity over seven baselines, and the authors note that generic benchmarks like VBench may penalize fixed-camera artifacts and reward drift-induced optical flow as “dynamic degree,” calling for new task-specific evaluation metrics.