Deep generative models, especially diffusion architectures, have transformed image generation; however, they are challenging to control and optimize for specific goals without expensive retraining. Embedding Space Exploration, especially with Evolutionary Algorithms (EAs), has been shown to be a promising method for optimizing image generation, particularly within Diffusion Models. Therefore, in this work, we study the performance of an evolutionary optimization method, namely Separable Covariance Matrix Adaptation Evolution Strategy (sep-CMA-ES), against the widely adopted Adaptive Moment Estimation (Adam), applied to Stable Diffusion XL Turbo's prompt embedding vector. The evaluation of images combines the LAION Aesthetic Predictor V2 with CLIPScore into a weighted fitness function, allowing flexible trade-offs between visual appeal and adherence to prompts. Experiments on a subset of the Parti Prompts (P2) dataset showcase that sep-CMA-ES consistently yields superior improvements in aesthetic and alignment metrics in comparison to Adam. Results indicate that the evolutionary method provides efficient, gradient-free optimization for diffusion models, enhancing controllability without the need for fine-tuning. This study emphasizes the potential of evolutionary methods for embedding space exploration of deep generative models and outlines future research directions.

Sparse autoencoders (SAEs) have become a core ingredient in the reverse engineering of large-language models (LLMs). For LLMs, they have been shown to decompose intermediate representations that often are not interpretable directly into sparse sums of interpretable features, facilitating better control and subsequent analysis. However, similar analyses and approaches have been lacking for text-to-image models. We investigated the possibility of using SAEs to learn interpretable features for a few-step text-to-image diffusion models, such as SDXL Turbo. To this end, we train SAEs on the updates performed by transformer blocks within SDXL Turbo's denoising U-net. We find that their learned features are interpretable, causally influence the generation process, and reveal specialization among the blocks. In particular, we find one block that deals mainly with image composition, one that is mainly responsible for adding local details, and one for color, illumination, and style. Therefore, our work is an important first step towards better understanding the internals of generative text-to-image models like SDXL Turbo and showcases the potential of features learned by SAEs for the visual domain. Code is available at https://github.com/surkovv/sdxl-unbox