# Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders

arXiv: https://arxiv.org/abs/2506.19708 · HTML: https://mlsanity.com/paper/2506.19708/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability.
- First posted on arXiv 24 Jun 2025.

## Placed in

- Mechanistic interpretability / [Interpreting vision and multimodal models](https://mlsanity.com/mechinterp/interpreting-vision-and-multimodal-models/)

## Built on by (most cited first, 7 of 7)

- [Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework](https://mlsanity.com/paper/2509.18127/) (2025)
- [Mechanistic Interpretability with Sparse Autoencoder Neural Operators](https://mlsanity.com/paper/2509.03738/) (2025)
- [DriftScope: Measuring The Hidden Effects of Diffusion Model Adaptation](https://mlsanity.com/paper/2607.00183/) (2026)
- [When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities](https://mlsanity.com/paper/2607.08605/) (2026)
- [Coverage-aware Semantic Representation Learning for Underrepresented Visual Concepts](https://mlsanity.com/paper/2602.16917/) (2026)
- [SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events](https://mlsanity.com/paper/2608.20117/) (2026)
- [From Interpretability Methods to Interpretable Models](https://mlsanity.com/paper/2609.05399/) (2026)

## Abstract (arXiv)

Despite their impressive performance, generative image models trained on large-scale datasets frequently fail to produce images with seemingly simple concepts -- e.g., human hands or objects appearing in groups of four -- that are reasonably expected to appear in the training data. These failure modes have largely been documented anecdotally, leaving open the question of whether they reflect idiosyncratic anomalies or more structural limitations of these models. To address this, we introduce a systematic approach for identifying and characterizing "conceptual blindspots" -- concepts present in the training data but absent or misrepresented in a model's generations. Our method leverages sparse autoencoders (SAEs) to extract interpretable concept embeddings, enabling a quantitative comparison of concept prevalence between real and generated images. We train an archetypal SAE (RA-SAE) on DINOv2 features with 32,000 concepts -- the largest such SAE to date -- enabling fine-grained analysis of conceptual disparities. Applied to four popular generative models (Stable Diffusion 1.5/2.1, PixArt, and Kandinsky), our approach reveals specific suppressed blindspots (e.g., bird feeders, DVD discs, and whitespaces on documents) and exaggerated blindspots (e.g., wood background texture and palm trees). At the individual datapoint level, we further isolate memorization artifacts -- instances where models reproduce highly specific visual templates seen during training. Overall, we propose a theoretically grounded framework for systematically identifying conceptual blindspots in generative models by assessing their conceptual fidelity with respect to the underlying data-generating process.