Research map / Image generation / Editing and style transfer
Sparse Relaxed-Lasso Steering: Automatic Sparse Autoencoder Feature Selection for Precise Image Editing
Zongxin Liu, Xiaoyong Xue, Sun Weidi, Shengchao Qin, Lijun Zhang
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 |
|---|---|---|---|
| replace dense All features with Lasso sparse support selection (ablation) | SRLS · Reduced gamma sweep (Appendix H) | LLM harmonic mean: 3.08 -> 3.71 | Improved |
| add support-restricted ridge refitting after Lasso (ablation) | SRLS · Reduced gamma sweep (Appendix H) | LLM harmonic mean: 3.71 -> 3.77 | Improved |
| use mean pooling instead of max pooling for the global attribute shift (ablation) | SRLS · Reduced gamma sweep (Appendix H) | LLM harmonic mean | Improved |
| use more than 50 contrastive pairs per attribute (ablation) | SRLS · Reduced gamma sweep (Appendix H) | LLM harmonic mean | No clear effect |
| restrict support range to [5, 500] rather than unconstrained search (ablation) | SRLS · Reduced gamma sweep (Appendix H) | LLM harmonic mean | No clear effect |
| add Lasso feature selection versus heuristic SAEdit selection (all-token steering) | SAEdit · 33 attributes, all-token steering | LLM harmonic mean: 3.89 -> 3.97 | Improved |
| add Bayesian optimization strength selection to SRLS | SRLS · All-token steering | LLM harmonic mean: 3.97 -> 4.24 (+0.27) | Improved |
| add Bayesian optimization strength selection to SAEdit | SAEdit · All-token steering | LLM harmonic mean: 3.89 -> 4.18 (+0.29) | Improved |
| add Bayesian optimization strength selection to All | All · All-token steering | LLM harmonic mean: 3.32 -> 3.93 (+0.61) | Improved |
| use SRLS instead of All for feature selection (feature count) | All · 33 attributes | mean selected features: 65536 -> 307.39 | Improved |
| use SRLS instead of SAEdit for Professions subjects | SAEdit · Subject category: Professions | LLM harmonic mean: 4.13 -> 4.22 | Improved |
Closest papers by meaning
- Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering (NeurIPS 2026 · similarity 0.89)
- One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Sparse Autoencoders, Again? (ICML 2025 · similarity 0.86)
- Accelerating Text-to-Image Editing via Cache-Enabled Sparse Diffusion Inference (AAAI 2024 · similarity 0.86)
- Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.85)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 2024 · similarity 0.85)
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.
