Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation

Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Felix Krause, Björn Ommer · 2026

Read Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation by Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Felix Krause, Björn Ommer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-19141.txt.

Description

Schusterbauer and five colleagues explore patch noise scales: some regions denoise, others need refinement. Naive variation across tokens performs poorly, exposing training states absent at inference. They introduce a sampler capping patch information during training; patch timesteps beat standard baselines. A patch difficulty head enables samplers allocating compute where needed. With noise varying over space and time, Patch Forcing advances easier regions first to provide context for harder ones. Patch Forcing beats baselines on ImageNet and scales to text-guided synthesis.

Catalog record

CREATORJohannes Schusterbauer, Ming Gui, Pingchuan Ma, Felix Krause, Björn Ommer
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61010
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.19141

Browse the full library · About the Tower of Babel