Research map / Image generation / Autoregressive and latent generation
Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders
Maty Bohacek, Thomas Fel, Maneesh Agrawala, Ekdeep Singh Lubana
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compare generated vs real LAION images in SAE concept space across four models | SD 1.5, SD 2.1, PixArt, Kandinsky · Energy difference distribution over 32,000 RA-SAE concepts (10,000 LAION pairs) | energy-difference skewness: skewness -0.54 (SD 2.1), -0.40 (SD 1.5, PixArt), -0.23 (Kandinsky) | Mixed |
| compare SD 2.1 with SD 1.5 on concept blindspot profile | SD 2.1 vs SD 1.5 · Cross-model concept blindspot correlation | Pearson r: r = 0.82 | No clear effect |
| compare PixArt and Kandinsky with SD 1.5 on concept blindspot profile | PixArt, Kandinsky · Cross-model concept blindspot correlation | Pearson r: r = 0.41 (PixArt), r = 0.46 (Kandinsky) | Mixed |
| apply DPO fine-tuning to SD 1.5 | SD 1.5 · Datapoint-wise concept energy difference | L2 norm of concept energy difference: lower median and narrower spread (no numbers given) | Improved |
| higher natural frequency of a concept | SD 1.5/2.1, PixArt, Kandinsky · Concept frequency vs absolute energy difference | energy discrepancy: higher-frequency concepts show lower energy discrepancies (no numbers given) | Improved |
| use alpha-masked image instead of heatmap overlay for VLM concept description (ablation) | GPT-4o autointerpretation · Qualitative concept description | description correctness | Improved |
Closest papers by meaning
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.90)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.89)
- Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (ICML 2025 · similarity 0.87)
- Cascaded Sparse Autoencoders LearnMulti-Level Visual Concepts in Multimodal LLMs (NeurIPS 2026 · similarity 0.87)
- Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models (ICML 2025 · similarity 0.87)
- Salient Concept-Aware Generative Data Augmentation (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
