SharedSAE: a unified sparse dictionary across u-net blocks and denoising time for diffusion models
Concept representations in diffusion models drift across network depth and denoising time, yet existing sparse autoencoders offer only per-block or per-timestep dictionaries, leaving no single feature dictionary for localization and intervention in the network.
This article introduces SharedSAE, a novel approach to sparse autoencoders in diffusion models that provides a unified feature dictionary across U-Net blocks and denoising steps.
What happened
The article presents SharedSAE, a method that creates a unified sparse dictionary for diffusion models by addressing the issue of concept representation drift across network depth and denoising time. It introduces technical components like a block-specific input adapter and a temporal conditioning module to manage the challenges of shared feature representation across different U-Net blocks and denoising steps.
Why it matters
SharedSAE provides a unified representation space for diffusion models, enabling more effective concept analysis and intervention. This is crucial for improving model interpretability, controlling generated content, and enabling targeted modifications in diffusion processes.
- SharedSAE offers a unified sparse dictionary across U-Net blocks and denoising steps in diffusion models
- The method enables localized feature suppression and concept erasure
- It improves model interpretability and controllability in diffusion processes
- Concept representations shift across U-Net blocks during denoising
- The approach addresses challenges of temporal non-stationarity and residual distribution alignment
The bigger picture
The method allows for more precise control over diffusion model outputs through localized feature suppression and concept erasure. It advances the ability to analyze and modify generated content at specific stages of the denoising process, improving model transparency and controllability.
SharedSAE introduces a unified sparse dictionary for diffusion models that enables better control, analysis, and intervention in the generation process by addressing concept representation drift across network depth and denoising time.
Technical details
- Sparse autoencoder (SAE)
- U-Net attention blocks
- Low-rank input adapter
- Temporal conditioning module
- Shared timestep branch
- Block-timestep correction branch
- Denoising steps
- SDXL 1.0
- I2P detector
- NudeNet-positive generations
FAQ
What is SharedSAE?
SharedSAE is a novel sparse autoencoder that creates a unified feature dictionary across U-Net blocks and denoising steps in diffusion models.
How does SharedSAE address concept representation drift?
It uses a block-specific low-rank input adapter to align residual distributions and a temporal conditioning module to handle temporal non-stationarity, maintaining a coherent dictionary structure across the network and denoising process.
What are the benefits of SharedSAE?
It enables localized feature suppression, concept analysis, and erasure at any block and denoising step, improving model interpretability and controllability.
What is the significance of the I2P detector-positive subset reduction?
The reduction from 139 to 1 NudeNet-positive generations demonstrates the effectiveness of SharedSAE in controlling and modifying generated content.
How does SharedSAE handle temporal non-stationarity?
It employs a temporal conditioning module with a shared timestep branch and a block-timestep correction branch to absorb temporal non-stationarity across denoising steps.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.CV (CV) on Oct 14, 2026.
Comments
Sign in to join the discussion.