pith. sign in

arxiv: 2605.19739 · v2 · pith:F4UYJNAXnew · submitted 2026-05-19 · 💻 cs.CV

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

Yi Sun , Zhiqi Zhang , Xinhao Zhong , Yimin Zhou , Shuoyang Sun , Bin Chen , Shu-Tao Xia , Ke Xu This is my paper

Pith reviewed 2026-05-20 05:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords concept erasureflow matchingreward optimizationGRPOtext-to-imagemodel safetyadversarial robustness
0
0 comments X

The pith

FlowErase-RL reframes concept erasure in flow matching models as a reward optimization problem using a dynamic dual-path reward system.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that concept erasure can be achieved more effectively by optimizing rewards rather than through supervised fine-tuning or inference-time fixes. It introduces two reward paths: one that penalizes generation of the target concept and another that encourages fidelity in non-target areas. These paths are balanced dynamically based on performance during training. This leads to better erasure results on things like nudity and styles while keeping image quality high and resisting attempts to bypass the erasure. A sympathetic reader would care because it suggests a scalable way to control what AI image generators produce without needing lots of matched data.

Core claim

The central claim is that reformulating concept erasure as a GRPO-based reward optimization problem, with a dynamic dual-path reward mechanism that balances Concept Erasure and Non-target Space rewards via a performance-driven switching strategy, enables state-of-the-art performance in suppressing target concepts while preserving generative quality and semantic alignment in flow matching models.

What carries the argument

The dynamic dual-path reward mechanism that jointly optimizes a Concept Erasure reward to suppress target concepts and a Non-target Space reward to preserve generative fidelity, adaptively balanced by a performance-driven switching strategy.

If this is right

  • The method achieves state-of-the-art erasure performance on nudity, object, and artistic style tasks.
  • It maintains strong image quality and semantic alignment after erasure.
  • It shows robust resistance to adversarial attacks.
  • It scales effectively to multi-concept erasure scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This approach may extend to other types of generative models that use similar flow or diffusion processes.
  • Reducing reliance on supervised data could make safety measures easier to implement across different concepts.
  • Testing the method on more complex or abstract concepts could reveal additional strengths or limits.

Load-bearing premise

The performance-driven switching strategy between the Concept Erasure and Non-target Space rewards can stably optimize the model without explicit supervision and without the two paths conflicting in ways that degrade either erasure or fidelity.

What would settle it

Observing training instability where the switching causes either poor erasure of targets or degraded quality in non-target images would challenge the central claim.

Figures

Figures reproduced from arXiv: 2605.19739 by Bin Chen, Ke Xu, Shuoyang Sun, Shu-Tao Xia, Xinhao Zhong, Yimin Zhou, Yi Sun, Zhiqi Zhang.

Figure 1
Figure 1. Figure 1: Overview of FlowErase-RL. (a) illustrates the framework of our approach. We employ [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of Nudity erasure results in I2P dataset and under attacks. 5 Experiments 5.1 Experimental setup Baselines. We compare our method against four SOTA approaches applied to flow matching models, including two training-based methods ESD [36], EraseAnything [38] and one training-free method DVE [35]. Evaluation metrics. We evaluate FlowErase-RL on three CE tasks: nudity erasure, artist style erasure,… view at source ↗
Figure 3
Figure 3. Figure 3: , clearly demonstrates the erasure results of our method for target object concept. O rigin al M o difie d Chain Saw Tench English Springer Garbage Truck [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Additional results of I2P dataset. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional object erasure results in Figure 9 and Figure 10. In Figure 11, we compare our [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional results of adversarial attacks, including MMA, RAB,P4D and UnlearnDiff. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results of multiple concepts erasure. Our method can successfully erase target concepts in [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional results of erasing ’Van Gogh’. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of Van Gogh erasure results. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional erasure results of 5 object. For each concept, the images show both target [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional erasure results of the other 5 object. For each concept, the images show [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of images generated by different methods via MS-COCO dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of FlowErase-RL on MS-COCO Dataset for all three types of concept erasure [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
read the original 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 \emph{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 \textbf{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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces FlowErase-RL, a GRPO-based framework that reformulates concept erasure in flow matching models as a reward optimization problem. It proposes a dynamic dual-path reward mechanism that jointly optimizes a Concept Erasure (CE) reward to suppress target concepts and a Non-target Space (NS) reward to preserve generative fidelity, with the two paths adaptively balanced via a performance-driven switching strategy. The authors claim this yields state-of-the-art erasure performance on nudity, object, and artistic style tasks while maintaining image quality and semantic alignment, with additional robustness to adversarial attacks and scalability to multi-concept settings.

Significance. If the central empirical claims hold, the work offers a promising new paradigm for safe generation in flow matching models by replacing supervised fine-tuning with reinforcement learning and an adaptive dual-reward scheme. This could improve scalability and multi-concept handling compared to prior inference-time or SFT-based erasure methods. The paper is credited for the reformulation of erasure as GRPO reward optimization and for conducting experiments across multiple erasure categories.

major comments (2)
  1. [§3.2] §3.2 (Dynamic Dual-Path Reward Mechanism): The performance-driven switching strategy is asserted to enable stable optimization without explicit supervision and without the CE and NS paths conflicting, yet the manuscript supplies no formal convergence analysis, Lyapunov-style stability argument, or ablation that isolates the switching logic from fixed-weight baselines. In flow matching models with coupled continuous trajectories, even modest misalignment could accumulate across timesteps, so the absence of such analysis makes it difficult to confirm that observed gains derive from the adaptive mechanism rather than hyperparameter choices.
  2. [§4.3] §4.3 (Experimental Results): The claim of state-of-the-art erasure performance with maintained fidelity rests on the switching strategy, but without an ablation comparing adaptive switching to static reward weighting or reporting per-timestep reward conflict metrics, it remains unclear whether the dual-path mechanism is load-bearing for the reported improvements or whether simpler baselines would suffice.
minor comments (2)
  1. The abstract and introduction would benefit from a concise summary table of key metrics (e.g., erasure success rate, FID, CLIP score) against the strongest baselines to allow readers to assess the SOTA claim at a glance.
  2. [§3.2] Notation for the switching threshold and performance metric used to trigger path selection should be defined explicitly in the method section rather than left implicit in the algorithm description.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on our work. We respond to each major comment below and indicate the revisions we plan to make to the manuscript.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Dynamic Dual-Path Reward Mechanism): The performance-driven switching strategy is asserted to enable stable optimization without explicit supervision and without the CE and NS paths conflicting, yet the manuscript supplies no formal convergence analysis, Lyapunov-style stability argument, or ablation that isolates the switching logic from fixed-weight baselines. In flow matching models with coupled continuous trajectories, even modest misalignment could accumulate across timesteps, so the absence of such analysis makes it difficult to confirm that observed gains derive from the adaptive mechanism rather than hyperparameter choices.

    Authors: We appreciate this observation regarding the theoretical underpinnings of our dynamic dual-path reward mechanism. Providing a formal convergence analysis or a Lyapunov-style stability argument for the performance-driven switching in the setting of coupled continuous trajectories is a significant undertaking that we have not pursued in the current work, as our focus has been on empirical validation and practical effectiveness. To directly address the request for an ablation isolating the switching logic, we will add such an analysis in the revised manuscript, comparing the adaptive strategy against fixed-weight baselines. We believe this will help confirm that the gains stem from the adaptive balancing rather than specific hyperparameter selections. revision: partial

  2. Referee: [§4.3] §4.3 (Experimental Results): The claim of state-of-the-art erasure performance with maintained fidelity rests on the switching strategy, but without an ablation comparing adaptive switching to static reward weighting or reporting per-timestep reward conflict metrics, it remains unclear whether the dual-path mechanism is load-bearing for the reported improvements or whether simpler baselines would suffice.

    Authors: We agree that additional experimental evidence would strengthen the case for the dual-path mechanism. In the revised version of the manuscript, we will include an ablation study comparing the adaptive switching strategy to static reward weighting approaches. Furthermore, we will report per-timestep reward conflict metrics to illustrate how the performance-driven switching mitigates potential conflicts between the CE and NS rewards, thereby supporting that the mechanism is indeed load-bearing for the observed state-of-the-art results. revision: yes

standing simulated objections not resolved
  • Formal convergence analysis or Lyapunov-style stability argument for the performance-driven switching strategy.

Circularity Check

0 steps flagged

No significant circularity in the proposed RL reformulation or empirical claims

full rationale

The paper introduces FlowErase-RL as a GRPO-based reward optimization framework with a dynamic dual-path (CE and NS) mechanism balanced by a performance-driven switching strategy. All central claims of SOTA erasure performance, fidelity preservation, and robustness are supported by extensive experiments across nudity, object, and style tasks rather than by any self-referential equations, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the results to the method's own inputs by construction. The derivation chain is therefore self-contained and externally validated through empirical benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the effectiveness of GRPO applied to flow matching and on the stability of the adaptive reward balancing; these are domain assumptions rather than derived results. No free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption GRPO can be applied to optimize concept erasure in flow matching models without requiring precisely aligned supervised data.
    The method is built directly on this premise to avoid SFT limitations.

pith-pipeline@v0.9.0 · 5765 in / 1302 out tokens · 52233 ms · 2026-05-20T05:23:45.725393+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.