Research map / Image generation / Text-to-image diffusion
ForceForget: Reinforcement Concept Removal for Enhancing Safety in Text-to-Image Models
Dong Han, Yong Li
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 |
|---|---|---|---|
| raise unsafe-score weight in Safety Evaluator from -1 to -2 (ablation) | ForceForget · I2P (sexual) | NRR: 82.92 -> 97.96 | Improved |
| use balanced CER weights (1,1) instead of (2,1) (ablation) | ForceForget · I2P (sexual) | NRR: 81.95 -> 97.96 | Improved |
| use 4 safe-adapter tokens instead of 2 (ablation) | ForceForget · I2P (sexual) | NRR: 87.43 -> 97.96 | Improved |
| use 4 safe-adapter tokens instead of 8 (ablation) | ForceForget · I2P (sexual) | NRR: 83.67 -> 97.96 | Improved |
| use 4 safe-adapter tokens instead of 16 (ablation) | ForceForget · COCO-30k | CLIP score: 27.924 -> 30.786 | Improved |
| apply safe adapter to end tokens rather than beginning tokens (ablation) | ForceForget · COCO-30k | CLIP score: 18.850 -> 30.786 | Improved |
| apply adapter to 32 end tokens rather than 32 beginning tokens (ablation) | ForceForget · I2P (sexual) | NRR: 92.80 -> 83.35 | Hurt |
| add alignment reward to safety reward (ablation) | ForceForget · COCO-30k | CLIP score: 21.1603 -> 30.53 | Improved |
| add safety reward to alignment reward (ablation) | ForceForget · I2P (sexual) | NRR: 93.89 -> 97.96 | Improved |
| add safe adapter to DDPO fine-tuning (ablation) | ForceForget · I2P (sexual) | reward mean / erasing | Improved |
| combine BLIP caption and target condition in alignment reward (ablation) | ForceForget · I2P (sexual) | erase-preserve tradeoff | Improved |
| transfer LoRA trained on SD v1.4 to SD v1.5 | ForceForget · I2P (sexual) | NRR: 78.3 -> 99.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DuMo — DuMo: Dual Encoder Modulation Network for Precise Concept Erasure (AAAI 2025) · better on 4 of 4 shared results
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.93)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.89)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.89)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.88)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.88)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 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.
