Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models

Researchers introduce residualized temporal sparse autoencoders (SAEs) to interpret how text-to-image diffusion models generate images over time. By analyzing activation trajectories across the denoising process rather than static snapshots, the method captures interpretable features that go beyond simple linear predictability, enabling better understanding of model internals.

This research advances interpretability in generative AI by addressing a fundamental limitation in existing analysis methods. While previous sparse autoencoder approaches treated diffusion model activations as static representations or simple time-conditioned snapshots, this work recognizes that diffusion models inherently operate through temporal sequences—the iterative denoising process creates activation trajectories that contain rich temporal structure. The key innovation lies in residualizing these trajectories by fitting linear predictors between consecutive timesteps, then training SAEs on the residual (unpredicted) components. This isolates non-linear, structurally significant features that vanilla approaches would miss.

The research builds on the growing recognition that AI safety and alignment depend critically on understanding model internals. Sparse autoencoders have become a primary tool for decomposing neural activations into human-interpretable directions, but their application to dynamic processes like diffusion required methodological innovation. By mapping decoded latents back into activation space as feature trajectories, researchers can visualize and analyze how specific concepts evolve during image generation.

For the AI development community, improved interpretability tools directly support model auditing, steering, and safety research. The demonstrated ablation studies and steering experiments on Stable Diffusion 1.5 suggest practical applications for controlling model behavior and understanding failure modes. This framework potentially extends beyond diffusion models to other sequential generative processes, broadening its utility for mechanistic interpretability research.

The work represents incremental but meaningful progress in a critical research direction. Future applications may include automated feature discovery, adversarial robustness analysis, and better control mechanisms for generative systems.

• →Residualized temporal SAEs capture diffusion activation structure missed by previous static or time-conditioned analysis methods
• →The approach decomposes full denoising trajectories into initial activations plus residual non-linear components for sparse encoding
• →Feature analysis becomes temporal, allowing researchers to track how specific concepts emerge throughout image generation
• →Successful ablation and steering experiments demonstrate practical applications for model control and interpretation
• →Framework generalizes beyond diffusion to any sequential neural process with learnable linear dynamics