Training-Free Adaptive Sampling Scheduler for Discrete-Time Consistency Distillation

Published in International Conference on Artificial Neural Networks (ICANN), 2026

Discrete-time consistency distillation methods each ship a timestep schedule welded to their own training framework and step budget, so they degrade as soon as the scheduler or the step count changes at inference time. We show that the schedule, not the student, is the bottleneck.

Method. An SNR-derived Importance measure scores each timestep by the normalised inverse rate of change of log-SNR; it peaks in the middle of the trajectory, where structure actually forms. Candidates above a threshold replace uniform-grid steps, so the step count never changes — only its spacing. A \(\gamma\)-I sampler scales noise re-injection by the local Importance instead of a constant, and a tanh soft clamp with channel and global mean re-centring removes the colour drift that classifier-free guidance compounds at low step counts.

Results. Across LCM, PCM, TCD and TDD on both SDXL (1024×1024) and SD v1.5 (512×512), every baseline improves on FID, CLIP Score and Inception Score at 2, 4 and 8 sampling steps. The gains concentrate where few-step sampling is weakest: at 2 steps FID falls 82–86% for PCM and TCD. The method is training-free, adds no parameters, and costs negligible runtime.

Read the full illustrated summary — equations, the complete results table and every figure, formatted for reading on a phone.

Download the conference poster (A0, PDF)

Publisher page (Springer)

Supported by JSPS KAKENHI JP25K15165 and the Kayamori Foundation of Informational Science Advancement.

Recommended citation: Qi Wang and Jinjia Zhou. (2026). "Training-Free Adaptive Sampling Scheduler for Discrete-Time Consistency Distillation." International Conference on Artificial Neural Networks (ICANN).
Publisher Page | Download Bibtex