Research map / Image generation / Editing and style transfer
Towards Safe Concept Transfer of Multi-Modal Diffusion via Causal Representation Editing
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 Causal Representation Editing (CRE), a training-free inference-time intervention that projects out unsafe-concept steering vectors from cross-attention in diffusion U-Nets during causally identified timesteps, to block unsafe object/style transfer in Kosmos-G.
Problem
Preventing multi-modal (vision-language-to-image) diffusion models from transferring unsafe or copyrighted concepts (objects or artistic styles) from user-supplied reference images, which text-based safe-generation methods cannot handle.
Approach
Extends LLM representation editing to diffusion models by building steering vectors from unsafe-concept example images, projecting them out of cross-attention outputs, gating edits with a per-concept discriminator, and restricting edits to a causal timestep interval found by an "assess-with-exclusion" search.
Key results
- On ImageNet object transfer, CRE reduces average top-1 transfer accuracy to 0% versus 36.26% for Kosmos-G, 22.48% for SLD and 1.66% for ProtoRe.
- On style transfer across four styles, CRE averages 2.12% versus 32.59% for Kosmos-G, 24.93% for SLD and 13.97% for ProtoRe.
- Editing overhead is about 0.2 seconds per image (226 s to 246 s for 100 images) on Kosmos-G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CRE instead of ProtoRe for unsafe object transfer | Kosmos-G · Safe object transfer (10 ImageNet classes) | Top-1 accuracy (%): 1.66 -> 0 | Helped |
| use negative-prompt Kosmos-G-Neg instead of unmodified Kosmos-G | Kosmos-G · Safe object transfer (10 ImageNet classes) | Top-1 accuracy (%): 36.26 -> 36.76 | Hurt |
| apply CRE instead of ProtoRe for unsafe style transfer | Kosmos-G · Safe style transfer (four styles; average) | Top-1 accuracy (%): 13.9715 -> 2.1202 | Helped |
| govern multiple unsafe styles instead of one | CRE · Multiple-style governance | Top-1 accuracy difference (%): +0.0854 | No clear effect |
| govern multiple unsafe styles instead of one | SLD · Multiple-style governance | Top-1 accuracy difference (%): +1.9022 | Hurt |
| govern multiple unsafe styles instead of one | ProtoRe · Multiple-style governance | Top-1 accuracy difference (%): +2.0728 | Hurt |
| apply CRE instead of ProtoRe for I2P safe generation | Kosmos-G · I2P (average) | Score: 0.09 -> 0.06 | Helped |
| add CRE inference editing | Kosmos-G · Image generation runtime (100 images) | Time (seconds): 226 seconds -> 246 seconds | Hurt |
| add projection-based representation editing (ablation) | CRE · Safe object and style generation | Image quality | Helped |
Surprising findings
- Kosmos-G-Neg, which appends negative prompts like 'without Van Gogh style', increases the probability of generating the unsafe object, suggesting the MLLM encoder fails to interpret negation.
Limitations
- Effectiveness depends on accuracy of the unsafe-concept discriminator; poor discriminators edit safe prompts.
- Representation editing adds inference overhead compared with fine-tuned safe generation models.
- Evaluation relies on classifiers trained on self-constructed style data and a limited set of 10 objects and 4 styles.
Experimental setup
Datasets: ImageNet DreamBench COCO-30k I2P Style Dataset Final (self-constructed)
Benchmarks: ImageNet 10-class object transfer DreamBench style transfer I2P
Metrics: top-1 classification accuracy of transferred concept (lower I2P category generation rate
Baselines: Kosmos-G Kosmos-G-Neg SLD ProtoRe
Builds on: Inference-Time Intervention ActAdd Safe Latent Diffusion Kosmos-G causal tracing
Closest papers by meaning
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.90)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.89)
- Anti-Diffusion: Preventing Abuse of Modifications of Diffusion-Based Models (AAAI 2025 · similarity 0.89)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
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.