acceptodds
Withdrawn from ICLR 2027

FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in flow matching models have significantly improved text-to-image generation quality, but also introduce growing safety risks due to the generation of harmful or undesirable content. Existing concept erasure methods are either inference-time interventions with limited effectiveness or rely on supervised fine-tuning (SFT), which requires precisely aligned data and struggles with scalability and multi-concept settings. In this paper, we propose FlowErase-RL, the first GRPO-based framework for concept erasure in flow matching models. We reformulate concept erasure as a reward optimization problem and introduce a dynamic dual-path reward mechanism that jointly optimizes (i) a Concept Erasure (CE) reward to suppress target concepts and (ii) a Non-target Space (NS) reward to preserve generative fidelity. The two reward paths are adaptively balanced during training via a performance-driven switching strategy, enabling stable optimization without explicit supervision. Extensive experiments on nudity, object, and artistic style erasure demonstrate that our method achieves state-of-the-art erasure performance while maintaining strong image quality and semantic alignment. Moreover, it exhibits robust resistance to adversarial attacks and scales effectively to multi-concept scenarios. Our results establish a new paradigm for safe and controllable generation in flow matching models.

Primary area:
generative models
Keywords:
Text-to-image, Concept Erasure, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HMvgOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FlowErase-RL, a GRPO-based framework for concept erasure in flow matching models. The method introduces a dynamic dual-path reward mechanism that combines a Concept Erasure (CE) reward for suppressing target concepts with a Non-target Space (NS) reward for preserving unrelated generation capabilities. A performance-driven switching strategy adaptively balances the two objectives during training. Experiments on nudity, object, and artistic style erasure demonstrate promising erasure effectiveness, adversarial robustness, and generation quality.

Critical strengths
  1. The research direction is sound and meaningful. The paper explores reinforcement learning as an alternative to conventional supervised fine-tuning for concept erasure in flow matching models, addressing a relevant problem in generative model safety.

  2. The framework is intuitive. The dual-path reward mechanism explicitly considers both erasure effectiveness and non-target preservation, providing a straightforward approach to balancing these objectives.

Critical weaknesses
  1. The methodological novelty is limited. The proposed framework primarily applies the existing Flow-GRPO algorithm to concept erasure, while the dual-path reward mechanism is relatively intuitive. The methodological contribution beyond adapting existing RL techniques and reward designs to this task remains somewhat limited.

  2. Justification for CLIP-based rewards is insufficient. The reward functions rely heavily on CLIP similarity, which does not always faithfully reflect visual or semantic similarity. Visually similar images may receive relatively low CLIP scores, raising concerns about whether directly optimizing CLIP-based rewards can reliably achieve concept erasure and non-target preservation.

  3. Questions regarding the ablation results. In Table 6, introducing the dual-path mechanism dramatically reduces ASRₑ from 88.8% to 0.0%, while ASRₖ remains almost unchanged (97.9% vs. 97.8%). This is somewhat counterintuitive given that the NS reward is primarily designed for non-target preservation. Furthermore, Figure 4 shows substantial sensitivity to ρmax, with ASRₑ increasing from 4.2% to 99.4% as ρmax changes from 0.3 to 0.4. These results appear somewhat unusual and warrant further explanation.

  4. The paper is not well written. The paper contains numerous inconsistencies that affect its overall clarity and presentation. For example, the metric directions for FID and CLIP are reversed in Table 6 and the reference to "table 4" in the ablation on ρmax is incorrect.

Questions for authors
  1. Could the authors further clarify the methodological novelty of FlowErase-RL beyond applying Flow-GRPO with relatively straightforward reward designs?

  2. Could the authors discuss the limitations of CLIP similarity as a reward signal and its potential discrepancies with actual visual similarity and concept preservation?

  3. Could the authors explain the unexpected results in Table 6, particularly the dramatic improvement in ASRₑ without a corresponding improvement in ASRₖ? Additionally, could they clarify the strong hyperparameter sensitivity observed in Figure 4?

Reviewer_u5X5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FlowErase-RL, which applies Flow-GRPO to erase concepts from FLUX.1-schnell. The method uses two rewards: an erasure reward, NSFW detector + CLIP for nudity, CLIP alone for objects and styles and a preservation reward computed on a retain set. Each training batch samples one of the two rewards with probability, updated per epoch according to the erasure reward signal. Evaluations cover nudity removal, erasure of ten ImageNet object categories, and art style removal.

Critical strengths

S1. Strong nudity erasure on FLUX. On I2P, nudity ASR drops to 8.80% vs 33.39% for the best baseline (DVE), and exposed parts go from 602 to 53 (Tables 1–2). It also has the lowest average ASR under the four attacks (Table 3).

S2. No paired target images needed. The erasure only needs prompt pairs and a scorer, so it is easy to apply. Table 12 even suggests the detector can be dropped for nudity.

S3. Several concept types. The paper tests nudity, objects and an artist style, and also checks adversarial attacks, which is good to see.

Critical weaknesses

W1. Limited novelty, many missing extensive amount of citations. RL- and preference-based erasure (Critic-Guided Reinforcement Unlearning, DUO, EraseFlow) and alternating forget/retain training (SCRUB, SalUn, AdvUnlearn) already exist, but none is cited, so the "first GRPO-based method" claim is too strong.

W2. Evaluation uses the same CLIP as training. The preservation reward optimizes CLIP on MS-COCO-like prompts and is then evaluated by CLIP on MS-COCO (31.33 → 32.57), so a recent strong model as independent judge (e.g. GPT, Gemini, Claude or Qwen-VL) is needed. More ablation studies on different models would be necessary.

W3. Limited amount of models. All main results use FLUX.1-schnell, and on SD v1.4 FID gets much worse (34.86 → 56.38, Table 5), so other flow models (e.g. SD3.5, Qwen-Image) are needed for the general claim.

W4. The switching rule is not tested. It is never compared with a fixed ratio or the summed reward, and erasure fails completely at ρmax⁡=0.4\rho_{\max}=0.4 (Fig. 4), although the paper says no concept-specific tuning is needed.

W5. Baselines and attacks are unclear. Only three baselines are run, the ESD port to FLUX is not described, and it is not clear if the attack prompts were optimized for each erased model.

W6. Some claims go beyond the results. Multi-concept scalability is only shown with example images (Fig. 9), and §1 says prior methods fail on FLUX, but two baselines (EraseAnything, DVE) are made for flow models.

Questions for authors
  1. How is the method different from SCRUB and EraseFlow, and can EraseFlow on FLUX be a baseline?
  2. How does the adaptive switching compare with a fixed ratio (e.g. ρ=0.15\rho = 0.15 or 0.30.3)?
  3. On what scale is the CLIP score, and how does ρt\rho_t change for objects and styles, given the shared thresholds 0.7 and 0.4?
  4. For objects and styles, what is c⊤c^\top?
  5. Can you evaluate with a recent VLM judge, and show results on another flow model?
  6. Were the attack prompts optimized against each erased FLUX model?
Reviewer_ZVieOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents the FlowErase-RL, a RL based method for concept erasure for text to image flow matching model. It treats erasure as reward maximization using Flow-GRPO. Experiments cover several Nudity benchmarks (including the adversarial ones), 10 objects and only one style. The paper uses the three baselines ESD, EraseAnything and DVE to compare against the proposed approach.

Critical strengths
  • Extends the Flow GRPo for concept erasure and proves that RL methods can be used for the concept erasure.
  • Some experiments are extensive, especially on nudity, and improves the performance against several baselines.
  • The paper is well written and easy to follow.
Critical weaknesses
  • Reward model is the same as the quantitative metric. CLIP is the main reward for objects and styles, part of the nudity reward, and part of the NS reward, yet CLIP score is the main fidelity metric.
  • Flux.1 is an old model and may not be enough to reproduce the latest models such as Flux.2, Minimax H2, etc.
  • Paper is missing another relevant paper as the potential baseline that also presents the RL-like approach, EraseFlow [1].
  • Additionally, the ablations don’t isolate the contributions.
    • The dual-path ablation (Table 6) covers a single concept (gas pump).
    • There are no performance reported across multiple seeds and variations.
  • Multi-concept scaling is claimed in the abstract but supported only by a qualitative figure (Fig. 9), with no quantitative results.
  • Artist style is tested on Van Gogh only, and other artists’ styles are not evaluated on.

[1] Kusumba, Naga Sai Abhiram, Maitreya Patel, Kyle Min, Changhoon Kim, Chitta Baral, and Yezhou Yang. "Eraseflow: Learning concept erasure policies via gflownet-driven alignment." Advances in Neural Information Processing Systems 38 (2026): 70876-70911.

Questions for authors

Please look at the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.