Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders

Matyas Bohacek, Thomas Fel, Maneesh Agrawala, Ekdeep Singh Lubana

Abstract

Generative image models often struggle with specific concepts, such as human anatomy or object counts, yet these failures are typically documented only through anecdotal evidence. This paper presents a systematic, unsupervised framework utilizing Sparse Autoencoders (SAEs) to identify "conceptual blindspots"—features in training data that models either suppress (under-generate) or exaggerate (over-generate). By mapping pixels to interpretable latent dimensions, the authors provide a quantitative comparison of concept distributions between natural and synthetic datasets across several state-of-the-art models.

Introduction

While generative models have reached impressive levels of realism, they exhibit persistent structural failures. Standard evaluation metrics like the Fréchet Inception Distance (FID) offer a global realism score but fail to pinpoint specific conceptual deficiencies. This research aims to move beyond qualitative "vibes-based" assessments toward a principled, quantitative measurement of how generative models deviate from real-world concept distributions.

Methodology

The proposed pipeline translates raw pixels into interpretable concepts through a multi-step process involving self-supervised learning and dictionary learning.

  1. Feature Extraction: High-level visual features are extracted from images using DINOv2.
  2. Archetypal Sparse Autoencoders (RA-SAE): An SAE with 32,000 latent dimensions is trained to decompose these features. To ensure stability and interpretability, the authors use an "archetypal" variant where dictionary atoms are constrained as convex combinations of actual training data.
  3. Energy-Based Metric: A metric $\delta(k)$ measures the "energy" difference for a specific concept $k$ between real and generated images. Values near 0 indicate suppressed blindspots (concepts the model ignores), while values near 1 indicate exaggerated blindspots (concepts the model over-uses).
The two-stage pipeline for concept extraction and analysis.

Experimental Setup

The RA-SAE was trained on the ImageNet-1k training split (~1.28M images). Evaluation was conducted using 10,000 image-text pairs sampled from LAION-5B. The authors tested four major open-weights models: Stable Diffusion 1.5, Stable Diffusion 2.1, PixArt-α, and Kandinsky 2.1.

Results and Insights

The analysis reveals that blindspots are not random but often reside in the "long tail" of the data distribution. There is a strong correlation between a concept's rarity in natural data and the model's failure to generate it accurately.

  • Suppressed Concepts: Models consistently fail to generate items like bird feeders, glossy DVD discs, and specific document layouts.
  • Exaggerated Concepts: Conversely, models often over-rely on "safe" textures or patterns, such as wood grain backgrounds or palm trees, even when not explicitly prompted.
  • Model Variance: While Stable Diffusion 1.5 and 2.1 share many blindspots ($r=0.82$), different architectures like PixArt and Kandinsky exhibit distinct conceptual maps ($r \approx 0.4$), indicating that architectural choices significantly shift a model's internal representation.
A spectrum of conceptual blindspots from suppressed (e.g., bird feeders) to exaggerated (e.g., wood textures).

Discussion

Further investigation at the datapoint level suggests that apparent successes in generating difficult concepts are often the result of memorization—reproducing specific visual templates rather than understanding the underlying concept. Interestingly, models fine-tuned with Direct Preference Optimization (DPO) show better alignment with ground-truth distributions, suggesting that human feedback helps mitigate these conceptual biases.

Conclusion and Limitations

This framework provides a new lens for evaluating generative models beyond aggregate scores. However, the method is currently limited by the perceptual capabilities of the underlying DINOv2 model and the sample size of the evaluation set. Future work will explore higher-order relational blindspots and more complex logical compositions.

Summary updated

BibTeX
@inproceedings{bohacek2026uncovering,
  author = {Matyas Bohacek and Thomas Fel and Maneesh Agrawala and Ekdeep Singh Lubana},
  title = {{Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=2sNrnTTEcv},
}
Figure