HTTP/2 301 server: GitHub.com content-type: text/html location: https://hila-chefer.github.io/Conceptor/ x-github-request-id: 92D5:3655F2:8DD4A0:9F1D83:69527847 accept-ranges: bytes age: 0 date: Mon, 29 Dec 2025 12:47:04 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210082-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1767012425.671908,VS0,VE200 vary: Accept-Encoding x-fastly-request-id: 016ba94c6c2adccf75a58b7a3fb362c8a125f5a1 content-length: 162 HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Wed, 24 Jan 2024 01:15:03 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"65b06497-5af9" expires: Mon, 29 Dec 2025 12:57:05 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: 7A5D:3946E9:8C5E27:9DA70F:69527848 accept-ranges: bytes age: 0 date: Mon, 29 Dec 2025 12:47:05 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210082-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1767012425.901179,VS0,VE225 vary: Accept-Encoding x-fastly-request-id: 3483123f46c1a0d50f3c6dfe8d99ce5149c7c3d4 content-length: 5295

The Hidden Language of Diffusion Models

Hila Chefer^1,2  Oran Lang¹  Mor Geva³  Volodymyr Polosukhin¹ 
Assaf Shocher¹  Michal Irani^1,4  Inbar Mosseri¹  Lior Wolf²

¹Google Research  ²Tel Aviv University  ³Google DeepMind  ⁴Weizmann Institute

ICLR 2024

arXiv Code

Explaining diffusion models with Conceptor. Our goal is to understand how a textual concept is represented by the model.

Concept decomposition with Conceptor: Given a concept (e.g., a president), and a text-to-image model, Conceptor learns to decompose the concept into a small set of interpretable tokens, with the objective of reconstructing the images generated for the concept. The decomposition reveals interesting behaviors such as reliance on exemplars (e.g., "Obama", "Biden").

Single image decomposition with Conceptor: Given a single image from the concept, our method extracts the tokens in the decomposition that caused the generation of the image. For example, a snail is decomposed into a ladybug and winding.

Abstract

Text-to-image diffusion models have demonstrated an unparalleled ability to generate high-quality, diverse images from a textual prompt. However, the internal representations learned by these models remain an enigma. In this work, we present Conceptor, a novel method to interpret the internal representation of a textual concept by a diffusion model. This interpretation is obtained by decomposing the concept into a small set of human-interpretable textual elements. Applied over the state-of-the-art Stable Diffusion model, Conceptor reveals non-trivial structures in the representations of concepts. For example, we find surprising visual connections between concepts, that transcend their textual semantics. We additionally discover concepts that rely on mixtures of exemplars, biases, renowned artistic styles, or a simultaneous fusion of multiple meanings of the concept. Through a large battery of experiments, we demonstrate Conceptor's ability to provide meaningful, robust, and faithful decompositions for a wide variety of abstract, concrete, and complex textual concepts, while allowing to naturally connect each decomposition element to its corresponding visual impact on the generated images.

How does it work?

Given a text-to-image diffusion model (e.g., Stable Diffusion), and the concept of interest (e.g., a president), Conceptor learns to decompose the concept into a small set of human-interpretable tokens from the model's vocabulary. This is achieved by mimicking the training process of the model. (1) We extract a training set of 100 images from the model using the concept prompt. (2) A learned MLP network maps each word embedding w_i from the model's vocabulary to its coefficient f(w_i). (3) We calculate the learned pseudo-token w^*_N as a linear combination of the tokens in the vocabulary weighted by their learned coefficients. (4) We sample a random noise for each of the training images and noise the images accordingly. (5) Using the learned pseudo-token w^*_N, we apply the model's UNet to predict the added noise for each image. (6) We compute two loss functions. L_reconstruct encourages the token w^*_N to reconstruct the training images, and L_sparsity encourages the learned coefficients to be sparse. (7) In inference time, we only consider the top 50 tokens rated by their coefficients to reconstruct the concept images.

Concept Decomposition

Single Image Decomposition

Given a single image generated for the concept, Conceptor can decompose the image into its own set of corresponding tokens that cause the generation. We find that the model learns to rely on non-trivial, semantic connections between concepts. For example, a snake is decomposed to a hose and a gecko such that the shape of its body is borrowed from the hose, and its head and skin texture are borrowed from the gecko.

@article{chefer2023hidden,
        title={The Hidden Language of Diffusion Models},
        author={Chefer, Hila and Lang, Oran and Geva, Mor and Polosukhin, Volodymyr and Shocher, Assaf and Irani, Michal and Mosseri, Inbar and Wolf, Lior},
        journal={arXiv preprint arXiv:2306.00966},
        year={2023}
}