Research map / Probabilistic and scientific modeling / Causal discovery
Beyond the Black Box: Identifiable Interpretation and Control in Generative Models via Causal Minimality
Lingjing Kong, Shaoan Xie, Guangyi Chen, Yuewen Sun, Xiangchen Song, Kun 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 |
|---|---|---|---|
| hierarchical SAE feature steering for nudity removal | SD 1.4 · I2P | nudity detection rate (%): 17.8 -> 0.25 | Improved |
| hierarchical SAE feature steering for nudity removal | SD 1.4 · P4D | nudity detection rate (%): 98.70 -> 0.66 | Improved |
| hierarchical steering vs best baseline MACE on Ring-A-Bell | SD 1.4 · Ring-A-Bell avg | nudity detection rate (%): 0.70 -> 1.05 | Hurt |
| hierarchical steering vs baselines on UnlearnDiffATK | SD 1.4 · UnlearnDiffATK | nudity detection rate (%): 0.70 -> 2.11 | Hurt |
| apply nudity steering to normal COCO prompts | SD 1.4 · COCO 10K | FID: 16.71 -> 17.02 | No clear effect |
| add hierarchical timestep modelling to SAE steering (ablation) | SD 1.4 · Add mountains | CLIP-I: 0.83 -> 0.91 | Improved |
| add hierarchical timestep modelling to SAE steering (ablation) | SD 1.4 · Add tabby pattern | CLIP-I: 0.83 -> 0.93 | Improved |
| add hierarchical modelling to SAE steering (ablation) | SD 1.4 · Replace rock w/ stump | CLIP-T: 0.29 -> 0.31 | No clear effect |
| higher noise timestep SAE deactivation (899 vs 100) (ablation) | SD 1.4 · SAE-deactivated generation | LPIPS: 0.002 -> 0.220 | Improved |
| use higher timestep SAE features (ablation) | SD 1.4 · spatial activation spread | top-1 spread: 0.27 -> 0.53 | Improved |
| remove sparsity constraint on SAE (ablation) | SD 1.4 SAE · concept graph (qualitative) | interpretability | Hurt |
| excessive sparsity in SAE (too small K) (ablation) | SD 1.4 SAE · concept graph (qualitative) | explanatory power | Hurt |
Closest papers by meaning
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.87)
- Learning Discrete Concepts in Latent Hierarchical Models (NeurIPS 2024 · similarity 0.87)
- Toward Causal Generative Modeling: From Representation to Generation (AAAI 2025 · similarity 0.87)
- Object centric architectures enable efficient causal representation learning (ICLR 2024 · similarity 0.87)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.86)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.86)
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.
