The model at low noise already knows what the model at high noise is struggling to learn.
Text-to-image diffusion models generate images by iteratively removing noise. At each timestep, the model predicts how to denoise — but the text-image alignment it's optimizing for is only implicitly supervised by the reconstruction loss. The connection between “remove noise correctly” and “match the text prompt precisely” is indirect.
Cross-Timestep Self-Calibration (arXiv:2603.20741): text-image alignment becomes progressively harder as the noise level increases. At low noise, the image is nearly clean and the model can reliably align it with the prompt. At high noise, the image is mostly noise and the alignment signal is buried. CTCal uses the reliable alignment from low-noise timesteps to explicitly supervise learning at high-noise timesteps. The easier version of the problem teaches the harder version.
This is self-distillation across difficulty levels. No external teacher, no additional training data — the model's own low-noise predictions provide the supervision signal for its high-noise predictions. The method is model-agnostic (works on both diffusion-based and flow-based architectures) and accepted to CVPR 2026.
The structural insight: in any system that processes the same input at multiple difficulty levels, the easy instances contain information about the hard ones. The model doesn't need an external alignment signal — it already produces one at low noise. The problem was that this signal wasn't being propagated to the regime where it's most needed. CTCal is a wiring fix, not a capability fix: connect the easy outputs to the hard inputs.