Research map / Image generation / Autoregressive and latent generation
Towards Test-Time Refusals via Concept Negation
Peiran Dong, Song Guo, Junxiao Wang, Bingjie WANG, Jiewei Zhang, Ziming Liu
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 |
|---|---|---|---|
| PROTORE vs plain Stable Diffusion | Stable Diffusion v2.1 · Imagenette erased class | top-1 accuracy of erased class (%): 83.84 -> 9.54 | Improved |
| PROTORE vs negative-prompt baseline | SD-Neg · Imagenette erased class | top-1 accuracy of erased class (%): 82.89 -> 9.54 | Improved |
| PROTORE vs safe latent diffusion | SLD · Imagenette erased class | top-1 accuracy of erased class (%): 52.96 -> 9.54 | Improved |
| PROTORE vs composable diffusion | CDM · Imagenette erased class | top-1 accuracy of erased class (%): 25.58 -> 9.54 | Improved |
| PROTORE (training-free) vs fine-tuned ESD on erased class | ESD · Imagenette erased class | top-1 accuracy of erased class (%): 12.6 -> 9.54 | Improved |
| PROTORE vs ESD on benign class preservation | ESD · Imagenette other classes | top-1 accuracy of other classes (%): 63.2 -> 74.28 | Improved |
| PROTORE vs unmodified SD on benign classes | Stable Diffusion v2.1 · Imagenette other classes | top-1 accuracy of other classes (%): 83.84 -> 74.28 | Hurt |
| PROTORE on large-area concept church | PROTORE · Imagenette church | top-1 accuracy of erased class (%): 98.0 -> 85.2 | No clear effect |
| PROTORE vs SD on inappropriate content | Stable Diffusion v2.1 · I2P | inappropriate probability: 0.39 -> 0.09 | Improved |
| PROTORE vs ESD and SLD on I2P | ESD / SLD · I2P | inappropriate probability: 0.19 -> 0.09 | Improved |
| PROTORE vs SD on image fidelity | Stable Diffusion v2.1 · COCO 30k | FID-30k: 14.50 -> 16.80 | Hurt |
| PROTORE vs SLD on image fidelity | SLD · COCO 30k | FID-30k: 16.90 -> 16.80 | Improved |
Closest papers by meaning
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.85)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.85)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.85)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.85)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.85)
- Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation (NeurIPS 2024 · similarity 0.84)
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.
