Noise Scheduling as Information-Guided Allocation in Diffusion Training

We introduce InfoNoise, an online adaptive noise schedule for diffusion training that reallocates optimization effort toward noise levels where denoising is most informative. Together with loss weighting, a schedule induces an effective allocation over denoising problems. InfoNoise makes this allocation data-adaptive by estimating a conditional-entropy-rate profile from denoising losses already computed during training, without auxiliary models or offline search. Through I--MMSE, this profile identifies where noisy observations most rapidly reduce uncertainty about the clean sample and provides the signal for adapting the training noise distribution. The method changes only this distribution, keeping the objective, loss weighting, and parameterization fixed. On image benchmarks, where schedules have been extensively tuned, InfoNoise matches or slightly exceeds strong fixed baselines and can reach the same final quality with fewer updates. On shifted representation, sequence, and modality settings, including DNA and language generation, where inherited schedules transfer poorly, InfoNoise improves baseline performance with up to $3\times$ less training compute. These results establish the conditional-entropy-rate profile as the data-dependent target for noise schedule design and make online adaptation a practical alternative to manual schedule search.