Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning

Details

Standard video diffusion and flow-matching systems rely on noise-level conditioning: the model is told at every denoising step how much noise is present. This creates a fixed sampling schedule that cannot adapt during inference. EqF removes this conditioning entirely, decoupling how the model learns to denoise from how sampling actually runs.\n\nThe result is a more modular architecture: inference algorithms can operate in a closed loop, using feedback from the generated sample to adapt the generation process. The paper shows this improves video quality and temporal consistency on challenging autoregressive video generation benchmarks versus noise-conditional baselines.\n\nThe key mechanism: without noise-level conditioning, the inference algorithm becomes data-dependent — it can adapt to what the model has generated so far rather than following a predetermined schedule. The authors provide extensive analysis of exactly why noise-level conditioning limits this adaptability.\n\nThe practical upside is inference-time flexibility: different sampling strategies become possible without retraining. The limitation is that results are specific to autoregressive video generation; broader applicability and human perceptual quality studies are not demonstrated in this paper.