Colorful-Noise: Training-Free Low-Frequency Noise Manipulation for Color-Based Conditional Image Generation 593

Nadav Z. Cohen, Ofir Abramovich, Ariel Shamir · 2026

Read Colorful-Noise: Training-Free Low-Frequency Noise Manipulation for Color-Based Conditional Image Generation 593 by Nadav Z. Cohen, Ofir Abramovich, Ariel Shamir (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-2605-00548.txt.

Description

Nadav Z. Cohen, Ofir Abramovich, and Ariel Shamir find a simple lever for controlling the broad look of images made by text-to-image diffusion models. These models turn white noise into a picture step by step. White noise has no structure, which gives variety but leaves global attributes like layout and color palette hard to steer or predict. The authors investigate which parts of the noise actually decide what.

Although all frequencies in the noise carry similar energy, the low-frequency parts mainly fix the image's global structure and color composition, while high frequencies fill in fine detail. So they manipulate only the low-frequency noise, using low-frequency cues taken from a guide image. The training-free method reconstructs those broad cues in the output while letting details emerge freely, preserving variety between samples. It adds minimal overhead, offering direct control over an image's structure and coloring.

Catalog record

CREATORNadav Z. Cohen, Ofir Abramovich, Ariel Shamir
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.00548

Browse the full library · About the Tower of Babel