Color encoding in Latent Space of Stable Diffusion Models

TL;DR

This paper investigates how color information is represented in the latent space of Stable Diffusion models, revealing an interpretable, opponent-axis encoding that enhances understanding and potential editing of generative models.

Contribution

It provides a systematic analysis showing that color is encoded along opponent axes in specific latent channels, offering new insights into the internal structure of diffusion models.

Findings

Color encoded along opponent axes in latent channels c_3 and c_4

Intensity and shape primarily represented in channels c_1 and c_2

Latent space exhibits an interpretable, efficient coding structure

Abstract

Recent advances in diffusion-based generative models have achieved remarkable visual fidelity, yet a detailed understanding of how specific perceptual attributes - such as color and shape - are internally represented remains limited. This work explores how color is encoded in a generative model through a systematic analysis of the latent representations in Stable Diffusion. Through controlled synthetic datasets, principal component analysis (PCA) and similarity metrics, we reveal that color information is encoded along circular, opponent axes predominantly captured in latent channels c_3 and c_4, whereas intensity and shape are primarily represented in channels c_1 and c_2. Our findings indicate that the latent space of Stable Diffusion exhibits an interpretable structure aligned with a efficient coding representation. These insights provide a foundation for future work in model understanding, editing applications, and the design of more disentangled generative frameworks.