Research map / Image generation / Text-to-image diffusion
EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
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 EraseFlow, a reward-free GFlowNet trajectory-balance method that fine-tunes Stable Diffusion v1.4 (LoRA) to erase concepts (nudity, artistic styles, logos) by aligning target-prompt denoising trajectories to anchor-prompt trajectories.
Problem
Concept erasure from text-to-image diffusion models is brittle: existing methods either degrade image quality, depend on adversarial losses or reward models, or require expensive retraining, and they treat all denoising steps equally.
Approach
Casts concept unlearning as trajectory matching with a GFlowNet trajectory-balance objective and a constant reward on anchor-prompt trajectories, removing the need for reward models and adversarial training while training over full denoising trajectories.
Key results
- On UDAtk nudity red-teaming, EraseFlow reaches ASR 33.89 vs 64.40 for the best non-adversarial baseline DUO; EraseFlow + AdvUnlearn reaches 1.42.
- Trains in about 2.8 minutes on a single A100 versus 225 minutes for R.A.C.E and 1440 minutes for AdvUnlearn.
- On fine-grained erasure, EraseFlow achieves the best total score (76.01 vs 71.32 for DUO) while concept score is 82.24 vs 93.97 for ESD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EraseFlow instead of DUO | Stable Diffusion v1.4 concept erasure · UDAtk nudity | ASR: 64.40 -> 33.89 | Helped |
| Combine EraseFlow with AdvUnlearn | Stable Diffusion v1.4 concept erasure · UDAtk nudity | ASR: 16.94 -> 1.42 | Helped |
| Combine EraseFlow with SAFREE | Stable Diffusion v1.4 concept erasure · UDAtk artistic style | similarity: 70.03 -> 62.88 | Helped |
| Use EraseFlow instead of DUO | Stable Diffusion v1.4 concept erasure · Fine-grained concepts | total score: 71.32 -> 76.01 | Helped |
| Use a close-to-target anchor prompt instead of the Fully Dressed anchor (ablation) | EraseFlow · Nudity I2P | ASR: 2.80 -> 44.30 | Hurt |
| Use a random unrelated anchor prompt instead of the Fully Dressed anchor (ablation) | EraseFlow · Nudity I2P | ASR: 2.80 -> 51.40 | Hurt |
| Use an empty-prompt anchor instead of the Fully Dressed anchor (ablation) | EraseFlow · Nudity I2P | ASR: 2.80 -> 9.80 | Hurt |
| Set logβ in the range [2,3] rather than small values (logβ ≤ 1) (ablation) | EraseFlow · Nudity erasure | erasure performance: 96% | Helped |
| Increase STOP_SAMPLING toward epoch 20 (ablation) | EraseFlow · Nudity erasure | performance | Helped |
| Use a balanced timestep mix (10 random steps from the first 40 and the last 10 steps) instead of only early or uniformly random steps (ablation) | EraseFlow · Nudity erasure | robustness | Helped |
Surprising findings
- Constant-reward trajectory balance without any reward model suffices for erasure, and the authors report that the detailed-balance variant was unstable and collapsed.
- Semantically opposite or neutral anchor prompts (e.g., 'Fully Dressed') erased better than semantically close anchors such as 'a model wearing lingerie'.
Limitations
- Peak memory of 42 GB is high relative to some baselines (e.g., UCE 0.40 GB, MACE 11.40 GB).
- Proposition relies on strong assumptions (exact zero loss, non-degenerate kernel), and its proof sketch uses per-timestep independence that is not justified.
- The reported CLIP score is slightly below some baselines (25.67 vs 26.36 for DUO).
Experimental setup
Datasets: I2P Ring-a-Bell MMA-Diffusion UDAtk MSCOCO
Benchmarks: I2P Ring-a-Bell MMA-Diffusion UDAtk MSCOCO Gecko fine-grained evaluation
Metrics: attack success rate mean cosine similarity (CSD) concept score total score CLIP score FID training time
Baselines: ESD UCE MACE DUO SAFREE R.A.C.E AdvUnlearn
Builds on: GFlowNets trajectory balance ESD AdvUnlearn SAFREE
Closest papers by meaning
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.92)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.91)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.90)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.90)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.90)
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.