Research map / Image generation / Text-to-image diffusion
Unveiling Concept Attribution in Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAD, a first-order linear attribution method over diffusion model parameters that finds concept-positive components (for erasure via CAD-Erase) and negative components (for amplification via CAD-Amplify).
Problem
Interpreting which model components (parameters) of text-to-image diffusion models contribute to generating a concept such as objects, styles, or nudity, beyond causal tracing of knowledge-storing layers.
Approach
Approximate the counterfactual effect of knocking out each parameter with a first-order Taylor term w_i * dJ/dw_i, computed from one forward and backward pass, yielding signed attribution scores; positive components erase a concept when ablated and negative components amplify it.
Key results
- CAD-Erase reduces nudity detections to 43 total on I2P versus 798 for SD-1.4, with FID 12.440 and CLIP-Score 31.30 on COCO, comparable to the base model.
- CAD-Amplify raises cassette player accuracy from 7.2% to 27.6% and most other target classes above 90% on ResNet50 classification.
- On SD-XL, CAD-Erase cuts nudity detections by 87.30% on I2P and 98.63% on MMA, while changing CLIP-Score by under 0.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Ablate top positive components for the target object | CAD-Erase on Stable Diffusion-1.4 · ImageNette objects | target-class accuracy: For church and garbage truck, CAD-Erase reduces accuracy to no more than 3% | Helped |
| Ablate top positive components | CAD-Erase on Stable Diffusion-1.4 · I2P nudity prompts | nudity detections (Total): 798 -> 43 | Helped |
| Ablate top positive components | CAD-Erase on Stable Diffusion-1.4 · COCO image prompts | FID: 14.127 -> 12.440 | Helped |
| Ablate negative components | CAD-Amplify on Stable Diffusion-1.4 · ImageNette target classes | target-class accuracy: Cassette player: 7.20 -> 27.60; Chain saw: 69.00 -> 98.20; Church: 76.20 -> 93.80; Gas pump: 79.00 -> 94.60; Tench: 80.4 | Helped |
| Ablate negative components | CAD-Amplify on Stable Diffusion-1.4 · I2P nudity prompts | nudity detections (Total): 798 -> 991 | Helped |
| Ablate positive components in the FF module rather than attention or residual modules (ablation) | CAD-Erase · Object erasure (10 ImageNette classes) | target-class accuracy: FF 1.82; Attn1 5.28; Attn2 25.40; Residual 18.96 | Mixed |
| Rescale negative components by 2 (ablation) | CAD intervention · ImageNette target classes | target-class accuracy: 78.26 -> 3.46 | Helped |
| Rescale negative components by 1.5 (ablation) | CAD intervention · ImageNette target classes | target-class accuracy: 78.26 | No clear effect |
| Rescale negative components by 3 (ablation) | CAD intervention · ImageNette other classes | other-class accuracy: 78.26 -> 35.79 | Hurt |
| Ablate positive components | CAD-Erase on Stable Diffusion XL · Ring-A-Bell prompts | nudity detections (Total): 537 -> 14 | Helped |
| Ablate positive components | CAD-Erase on Stable Diffusion 3.5 · I2P prompts | nudity detections (Total): 514 -> 157 | Helped |
| Remove too many positive components (ablation) | CAD-Erase · Object erasure | other-class accuracy | Hurt |
Surprising findings
- Ablating negative components (identified by attribution) substantially increases the probability of generating the target concept, revealing components that suppress knowledge.
- Pruning positive parameters outside feed-forward layers (self-attention, cross-attention, residual) strongly affects other concepts, indicating entanglement.
Limitations
- Only analyzes individual parameters; layers or modules that affect multiple concepts are not studied.
- Lightweight erasure/amplification can slightly degrade other classes, indicating entangled knowledge.
- Attribution targets only the final generated image, not specific denoising stages or spatial locations.
Experimental setup
Datasets: ImageNette I2P COCO MMA Ring-A-Bell
Benchmarks: ImageNette I2P COCO MMA Ring-A-Bell
Metrics: accuracy on target and other classes NudeNet detection count CLIP-Score FID LPIPS attack success rate
Baselines: ConceptPrune ESD UCE RECE SD-1.4 SD-2.1
Builds on: causal tracing ESD UCE ConceptPrune
Closest papers by meaning
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.91)
- Localizing and Editing Knowledge In Text-to-Image Generative Models (ICLR 2024 · similarity 0.91)
- Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.90)
- Blink of an eye: a simple theory for feature localization in generative models (ICML 2025 · similarity 0.89)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.89)
- Localizing Knowledge in Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
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.