acceptodds
Under review as a conference paper at ICLR 2027

Mosaic: Compositional multi-concept erasure via vector field blending

Abstract

Concept erasure has emerged as a key research direction for ensuring safe and ethical image synthesis in Text-to-Image (T2I) models. While existing studies have explored concept erasure across multiple concepts, they typically assume only a single target concept per image—a limitation increasingly exposed by modern flow-based T2I models, which can generate complex scenes with multiple concepts simultaneously. To address this gap, we introduce compositional multi-concept erasure, a new task that aims to simultaneously remove multiple target concepts within a single scene. We propose CoME-Bench, a benchmark for evaluating compositional multi-concept erasure, which covers both intra- and cross-category scenarios. We further propose Mosaic, a novel framework for multi-concept erasure in flow-based T2I models, which exploits the spatial locality of target concepts in the vector field by dynamically constructing concept-specific masks and selectively blending them without additional optimization. Extensive experiments demonstrate that Mosaic effectively removes multiple target concepts in complex compositional scenes while preserving non-target contexts. Our codes and datasets will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Concept erasure

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kCK6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Mosaic, a method for compositional multi-concept erasure in text-to-image models. It also introduces CoME-Bench to evaluate the erasure of multiple target concepts in the same image. Mosaic uses the vector field differences between the original model and concept-erased models to construct spatial masks. These masks are then used to blend different vector fields during generation. Experiments on FLUX.1-dev show that Mosaic outperforms several baseline methods in different compositional settings.

Critical strengths
  1. CoME-Bench provides a useful evaluation setting for multi-concept erasure. Unlike existing benchmarks that evaluate target concepts separately, it considers multiple target concepts appearing in the same image.
  2. Mosaic combines concept-specific erasure models through spatial vector field blending. This is a reasonable way to reduce the interference between different target concepts without additional training for their combination.
  3. The experiments cover different concept combinations and several baseline methods. The ablation studies and experiments on overlapping concepts are also useful.
Critical weaknesses
  1. Table 11 shows that several object concepts have relatively low erasure success rates. For example, the success rate for Wolf is only 16.3%. Meanwhile, Table 8(d) shows that Mosaic performs worse than MACE in the three-object setting. Although Appendix F provides some analysis, the performance on difficult compositions remains limited.
  2. The main experiments are conducted only on FLUX.1-dev. CoME-Bench also uses flow-based T2I models during its prompt generation and verification process. Therefore, it is unclear how well Mosaic generalizes to other T2I models.
  3. Mosaic requires multiple forward passes during sampling. Table 15 reports an inference latency of 2.63× and 3.44× compared with the original model when erasing two and three concepts. This cost increases with the number of target concepts.
Questions for authors

Please see the weaknesses above. I am willing to keep a positive rating if these concerns are properly addressed.

3 comments

W.1) Performance on difficult compositions

We thank the reviewers for highlighting this observation. We acknowledge that Mosaic achieves lower ESR than MACE in certain challenging compositional settings, particularly when the number of target concepts increases. However, as shown in our quantitative results, MACE exhibits substantially lower Selective Alignment (SA) scores than Mosaic (e.g., 0.73 vs. 0.89 in the three-object setting and 0.75 vs. 0.90 in the one-character + two-object setting, respectively), indicating that MACE tends to interfere more with the generation of non-target concepts. This suggests that the higher ESR achieved by MACE in these settings may come at the cost of preserving unrelated scene content.

To further investigate this trade-off, we conduct additional evaluations of non-target (NT) region preservation in the three-object and one-character + two-object settings using LPIPS, SSIM, and two DINOv2-based similarity metrics: DINO and DINO-patch [1],[2]. Specifically, we use SAM3 to identify non-target regions and obtain corresponding spatial masks from images generated by the original FLUX model. We then apply the same masks to the original FLUX-generated images and the corresponding erased images to assess how faithfully each method preserves the original appearance and structure of non-target content. DINO measures the cosine similarity between features averaged over non-target patches, while DINO-patch averages the cosine similarities between spatially corresponding non-target patch features. Both DINOv2-based metrics use the same non-target mask derived from the original image.

MethodNT-LPIPS (↓)NT-SSIM (↑)DINO (↑)DINO patch (↑)
MACE0.60310.46470.70700.3670
Mosaic (Ours)0.57720.46610.73560.3953

These results indicate that Mosaic better preserves non-target regions than MACE, revealing a trade-off between erasure effectiveness and non-target preservation in challenging compositional settings. Nevertheless, the absolute performance of both methods remains limited on these difficult tasks. We therefore agree that improving erasure effectiveness in challenging cases while preserving non-target content is an important direction for further research.

[1] Caron et al., Emerging Properties in Self-Supervised Vision Transformers, ICCV 2021

[2] Ruiz et al., DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation, CVPR 2023

W.2) Generalization to other T2I models

Thank you for raising this concern. To examine generalization beyond Flux.1-dev, we conduct an additional experiment on Flux.2-klein-base-4B, which was not used in CoME-Bench construction. This model uses a different text encoder and VAE, resulting in different text conditioning and image latent representations. We therefore train new concept-specific erasure LoRAs for this model and apply Mosaic to compose them.

We evaluate the Mario + SpongeBob combination using the same 96 prompts and generation seed. The table below reports ESR, Selective Alignment (SA) and Aesthetic score.

Method (Model)ESR-Gemma(↑)SA(↑)Aesthetic(↑)
EraseAnything (Flux.2-klein-base-4B)19.7993.755.152
MACE (Flux.2-klein-base-4B)50.009.033.782
Mosaic (Flux.2-klein-base-4B)42.7192.555.164

As shown in the results above, Mosaic is also applicable to Flux.2, achieving effective concept erasure while preserving non-target content and image quality. EraseAnything exhibits a similar trend to that observed on Flux.1 as in Table 8, which is outperformed by our method. In contrast, the MACE fusion rule does not transfer successfully to Flux.2 under our experimental settings: it produces severely degraded images, making its high ESR misleading as an indicator of successful concept erasure. This degradation is reflected in its substantially lower SA and aesthetic scores.

W.3) Efficiency

We acknowledge that computational efficiency is one of the limitations of Mosaic, as demonstrated by our experimental results in Tables 6 and 15. We agree with the reviewer that reducing the computational overhead is an important direction for future work.

We thank Reviewer kCK6 for the constructive feedback. We have addressed the reviewer's concerns in our rebuttal and would appreciate any follow-up on whether our response resolves the issues raised or whether any concerns remain.

My main concern remains the limited erasure effectiveness in challenging compositional object erasing. The new preservation results help explain the advantage over MACe, but the absolute erasure success rates for some concepts and three-concept combinations remain relatively low. The computational overhead also remains a limitation, as acknowledged by the authors. I therefore decide to maintain my original rating.

Reviewer_vWK4OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies compositional multi-concept erasure: removing multiple concepts from one generated image while preserving non-target content. It introduces CoME-Bench, with approximately 31K verified prompts, and Mosaic, which spatially combines concept-specific erasure LoRAs using vector-field discrepancy masks. Experiments on FLUX.1-dev show improved aggregate erasure success, supported by ablations and analyses of localization and individual eraser quality.

Critical strengths
  1. Mosaic offers an intuitive, modular approach: vector-field discrepancies localize interventions, independently trained LoRAs are reused without combination-specific optimization, and the base-model output preserves regions outside the masks.

  2. CoME-Bench verifies target presence across multiple random seeds, reducing confusion between generation failures and successful erasure. Extracting non-target elements also supports preservation evaluation. Ablations examine mask construction, scaling, and blending duration, while additional analyses cover overlapping concepts, inference overhead, and LoRA quality. Aggregate gains under two VLM evaluators support the qualitative findings.

Critical weaknesses
  1. Baseline coverage is limited, particularly for methods developed for DiT and flow-based generators. The current comparisons leave the advantage over other relevant erasure methods insufficiently established.

  2. The broader practical significance of CoME-Bench is insufficiently established. The paper motivates the benchmark through improved compositional generation and a gap in existing evaluations, but provides limited evidence that the selected scenarios represent an important practical failure mode in concept erasure.

Questions for authors
  1. Can you add comparisons with MCE (Minimalist Concept Erasure in Generative Models, ICML 2025), DP (Closed-Form Concept Erasure via Double Projections, CVPR 2026), and BLOCK (Concept Removal for Frontier Image Generative Models, ICML 2026), which evaluate erasure on FLUX and provide public implementations? Comparisons using the same erasers with effective weight-update merging and uniform vector-field blending would also help isolate the contribution of spatial composition.

  2. What practical use cases or observed failure patterns establish the importance of CoME-Bench’s compositional scenarios? What consequential limitations of existing evaluations does this benchmark reveal?

  3. The preservation evaluation provides limited evidence for the specificity of Mosaic’s spatial intervention. SA primarily checks whether extracted non-target elements remain present, while aesthetic scores and human ratings assess overall image quality. These measures do not directly establish how faithfully non-target regions retain their original appearance and structure. Can you provide a more direct evaluation of preservation in non-target regions?

7 comments

W.1, Q.1) Additional comparisons with baselines

We thank the reviewer for highlighting these recent concept erasure methods and suggesting additional comparisons. Following the reviewer's suggestions, we conduct additional experiments with Double Projection (DP), BLOCK, and Minimalist Concept Erasure (MCE), as well as effective weight-update merging and uniform vector-field blending, to evaluate their effectiveness in simultaneously erasing two target concepts within a single image and provide a more comprehensive evaluation of Mosaic. While DP, BLOCK, and MCE primarily focus on erasing individual target concepts, Mosaic specifically addresses compositional multi-concept erasure, aiming to simultaneously remove multiple co-occurring target concepts from a single image while preserving non-target content. This distinction highlights the unique objective and contribution of our work.

Since CoME-Bench focuses on simultaneously erasing multiple target concepts appearing in a single image, we adapt the three existing methods to remove multiple target concepts within a single generation:

  • DP modifies text-conditioning weights by projecting target representations toward a safe subspace while preserving designated non-target concepts. For compositional erasure, we provide multiple target concepts and their corresponding surrogate concepts to the original multi-concept editing routine. The method sequentially updates the same weights, producing a single edited checkpoint for each concept pair.

  • BLOCK uses sparse transcoders to identify concept-related features in text-conditioning layers and removes target concepts by modifying the corresponding decoder weights. To adapt its single-concept implementation to our compositional multi-concept erasing, we sequentially apply the concept-specific decoder edits to the same shared transcoders. We use a fixed editing order; when both edits modify the same decoder row, the later edit overwrites the earlier one.

  • MCE learns component-wise masks to suppress target concepts while preserving unrelated content. For compositional erasure, we independently train a mask for each target concept, binarize the masks, and combine them through element-wise multiplication of their keep masks. Consequently, a component is suppressed if either concept-specific mask suppresses it, allowing multiple concepts to be erased without pair-specific retraining.

To further isolate the contribution of Mosaic's spatial vector-field composition, we implement two additional baselines using the same independently trained single-concept erasure LoRAs used in Mosaic:

  • Effective Weight-Update Merging (CtM): Compress-then-Merge combines multiple LoRA adapters by first projecting their weight updates into shared low-rank subspaces, then merging them in this compact space. The merged update is mapped back to produce a single rank-constrained LoRA adapter.

  • Uniform Vector-Field Blending: We independently activate each concept-specific LoRA and compute its vector field at the same latent, timestep, and prompt. We then take the arithmetic mean of these vector fields: vuniform=1N∑i=1Nvi,v_{\mathrm{uniform}}=\frac{1}{N}\sum_{i=1}^{N}v_i, where NN denotes the number of target concepts. Unlike Mosaic, this approach does not use spatial masks, discrepancy-based weighting, or an explicit base-model contribution.

Due to time constraints, we conduct these comparisons on a subset of CoME-Bench. Specifically, we select one concept from each of three different categories: character, object, and NSFW content. For the character and object categories, we exclude concepts with poorly performing single-concept erasure LoRAs and randomly select one concept from the remaining candidates. This results in three selected concepts—Buzz (character), Louis Vuitton Backpack (object), and Nudity (NSFW content). We evaluate their pairwise combinations, comprising a total of 252 prompts. Following the evaluation protocol in our paper, we measure the Erasure Success Rate (ESR) using Gemma4, Selective Alignment (SA), and Aesthetic score. The results are presented below.

MethodESR-Gemma(↑)SAAesthetic
DP75.7994.965.74
BLOCK11.110.403.53
MCE36.9093.285.92
CtM57.1496.005.93
Uniform Blending49.2195.985.96
Mosaic (Ours)79.3795.985.91

As shown in the table, Mosaic achieves the highest ESR (79.37%) among all compared methods, while maintaining SA and aesthetic scores comparable to those of the best-performing methods. These results demonstrate the effectiveness of Mosaic's spatial vector-field composition for simultaneous multi-concept erasure. In particular, we observed that applying BLOCK to compositional multi-concept erasure often resulted in severely distorted or degraded images, leading to substantially lower erasure performance.

W.2, Q.2) Significance of CoME Bench

We thank the reviewer for raising this important question. We would like to further clarify the practical motivation and significance of CoME-Bench.

In real-world text-to-image generation scenarios, multiple concepts that need to be erased may co-occur within a single image. For instance, an image may contain multiple copyrighted characters, branded objects, or concepts subject to content restrictions. In such cases, concept erasure methods should simultaneously remove all specified target concepts while preserving unrelated scene content (e.g., erasing both Mario and SpongeBob in an image, as exemplified in Fig. 1).

Existing benchmarks, such as UnlearnCanvas, HUB, EraseBench, SEE, and EMMA, do not explicitly focus on the simultaneous removal of multiple target concepts within a single generated image. CoME-Bench complements these benchmarks by directly evaluating this compositional erasure setting. Reviewers FdsB and kCK6 also recognized this compositional evaluation setting as a key strength of CoME-Bench.

To further demonstrate the necessity of compositional prompt verification, we conduct an additional experiment using prompts from existing benchmarks. The goal is to examine the possibility that compositional multi-concept erasure can be properly evaluated using existing single-concept benchmarks. Specifically, we select single-concept prompts for Mario, SpongeBob, and Angelina Jolie and concatenated pairs of prompts to construct compositional prompts. We consider three pairwise combinations, with 25 prompts per combination, resulting in 75 prompts in total. Following the prompt verification protocol used in CoME-Bench, we generate images using FLUX.1-dev and evaluate the presence of both target concepts using Qwen3. Only 17 out of 75 prompts (22.7%) successfully generate images containing both target concepts. This result demonstrates that simply combining existing single-concept prompts does not reliably produce images in which all specified concepts co-occur and thus cannot be used for multi-concept erasure benchmarking, underscoring the importance of explicit co-occurrence verification in constructing a compositional erasure benchmark, as in our benchmark.

Moreover, strong performance in single-concept erasure does not necessarily guarantee successful erasure when multiple target concepts appear together. Combining multiple erasure interventions may introduce interference between concept-specific modifications, resulting in incomplete target removal or unintended degradation of non-target content. Such failure modes cannot be fully captured by conventional single-target evaluations.

To address this gap, CoME-Bench evaluates the simultaneous removal of co-occurring target concepts across both intra-category and cross-category compositions. Our experimental results reveal substantial performance variations across different compositional settings, particularly in challenging three-concept compositions. These findings highlight the importance of evaluating concept erasure beyond isolated target concepts.

Therefore, CoME-Bench provides a complementary and practically relevant evaluation setting that measures the reliability of simultaneous multi-concept erasure, which cannot be fully characterized by existing single-target evaluations.

Q.3) Additional evaluation of preservation in non-target regions

We thank the reviewer for this valuable suggestion. To provide a more direct evaluation of non-target region preservation, we conduct additional experiments using LPIPS, SSIM and DINOv2 similarity.

Specifically, we use SAM3 to identify non-target regions in images generated by the original FLUX model and obtain corresponding spatial masks. We then apply the same masks to the original and erased images to evaluate how faithfully the non-target regions preserve their original appearance and structure. We measure perceptual similarity using LPIPS (↓), SSIM (↑) and feature similarity using DINOv2 [1],[2] (↑).

The results are presented below.

2 concepts

MethodNT-LPIPS (↓)NT-SSIM (↑)DINO (↑)DINO patch (↑)
EraseAnything0.58380.45140.81380.4164
MACE0.61070.43660.75950.3798
Mosaic (Ours)0.58090.45120.80750.4150

3 concepts

MethodNT-LPIPS (↓)NT-SSIM (↑)DINO (↑)DINO patch (↑)
EraseAnything0.58820.45530.74690.3945
MACE0.61110.45320.70700.3644
Mosaic (Ours)0.58720.45200.73340.3907

Mosaic achieves the best performance in NT-LPIPS, while EraseAnything performs best on the other preservation metrics, with only marginal differences between the two methods. In contrast, MACE exhibits substantially worse preservation performance. Similar trends are observed in the three-concept settings, except for NT-SSIM.

[1] Caron et al., Emerging Properties in Self-Supervised Vision Transformers, ICCV 2021

[2] Ruiz et al., DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation, CVPR 2023

Reviewer_FdsBOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Mosaic, a framework for erasing multiple concepts from an image generated by flow-based T2I methods. CoME-Bench, a new benchmark containing around 31,000 prompts for evaluating multi-concept erasure in different settings.

Mosaic uses concept-specific LoRA models and dynamically generated spatial masks to combine vector fields and erase multiple target concepts while preserving the non targeted parts of the image .

Experiments such as quantitative comparisons, ablation studies, and a user study are effective.

Critical strengths

Benchmark for compositional multi-concept erasure: CoME-Bench, evaluating the erasure of multiple target concepts appearing together in a single image. Unlike traditional benchmarks that mainly evaluate concepts separately, CoME includes intra-category and cross-category. The benchmark has around 31,000 prompts and provides a useful resource for studying more challenging concept-erasure settings.

An interesting approach using spatial vector field blending: Mosaic uses differences between the base model and concept-erased models to identify regions associated with each target concept. It then combines the corresponding vector fields to remove multiple concepts while preserving non-target content. This is an interesting way to reuse existing single-concept erasers without additional joint optimization.

Extensive experiments and useful ablation studies.

Critical weaknesses

Limited novelty and missing comparisons with related methods: Earlier works, such as Receler (ECCV 2024) and GLoCE (CVPR 2025), explored localized concept erasure and these were not compared with these methods

Unfair baseline comparisons:MACE and EraseAnything adapted to the FLUX architecture, but it may not preserve the original methods' effectiveness.

Missing comparisons with simpler blending strategies; paper did not provide comparisons with other methods, such as direct vector field averaging or blending without spatial masks.

Low performance on challenging multi-concept settings: Table 8, 10.83% ESR with Gemma for 3-object erasure, while MACE's 19.52%. same with the Table 9 showing weaker results for 1-character and 2-object compositions showing method effectivness when the no. of target concepts increases.

Limitations of the proposed mask assignment strategy: The masks are assigned using a static threshold of 0.5. As per the equation in section 4.3 the concept must contribute a large enough portion of the vector field difference. If a number of concepts are contributing similar amounts then none will meet the threshold. This can cause issues if there are overlapping concepts or 3 or more targets in the scene.

Questions for authors

Dependence on single-concept erasers, Computational efficiency missing

See Weakness

6 comments

W.1) Limited novelty and missing comparisons with related works

We thank the reviewer for highlighting Receler (ECCV 2024) and GLoCE (CVPR 2025). We’d like to note that although these methods also exploit spatial locality for concept erasure, their localization mechanisms and objectives differ from ours, as explained below.

Receler constructs target-specific spatial masks from cross-attention maps and uses them to regularize lightweight eraser adapters during training, encouraging the preservation of non-target regions. For compositional multi-concept erasure, Receler simply averages the outputs of independently trained erasers during inference.

GLoCE first collects image embeddings from generated images of target and mapping concepts and extracts concept-specific subspaces through PCA. It then constructs spatial gates based on these representations to selectively apply low-rank feature transformations. For multiple target concepts, GLoCE selects the module with the highest gate activation at each image embedding.

In contrast, Mosaic differs from Receler in how spatial localization is used for compositional multi-concept erasure. While Receler employs spatial masks during eraser training and simply averages the outputs of independently trained erasers during inference, Mosaic dynamically computes spatial masks from vector-field discrepancies at each sampling step and uses them to selectively blend concept-specific eraser outputs in their corresponding spatial regions. Furthermore, unlike GLoCE, Mosaic does not require additional concept-specific image generations or precomputed feature subspaces for localization. Instead, it directly exploits the differences between the vector fields of the base model and concept-specific erasers during inference.

We also experimentally compare our method against these two methods. Due to time constraints, we conduct comparison on a subset of CoME-Bench rather than the entire benchmark. Specifically, we select one concept from each of three different categories: character, object, and NSFW content. For the character and object categories, we exclude concepts with poorly performing single-concept erasure LoRAs and randomly select one concept from the remaining candidates. This results in three selected concepts—Buzz (character), Louis Vuitton Backpack (object), and Nudity (NSFW content). We evaluate their pairwise combinations, comprising a total of 252 prompts.

Following the evaluation protocol in our paper, we measure the Erasure Success Rate (ESR) using Gemma4. The results are presented in the table below.

MethodESR-Gemma(↑)SA(↑)Aesthetic(↑)
Receler71.8382.595.74
GLoCE69.4484.395.83
Mosaic (Ours)79.4095.985.91

As shown in the table, Mosaic achieves the highest ESR (79.40%), SA(95.98%), and Aesthtic (5.91), outperforming Receler (71.83%) and GLoCE (69.44%). These results demonstrate the effectiveness of Mosaic's inference-time spatial vector-field blending for compositional multi-concept erasure, highlighting its advantage over Receler and GLoCE in simultaneously removing multiple target concepts. We will add these results in the revised paper.

W.2) Unfair baseline comparisons

We thank the reviewer for raising concerns about the fairness of our baseline comparisons.

EraseAnything was originally developed for FLUX.1-dev and trains concept-specific LoRAs for concept erasure. For compositional multi-concept erasure, EraseAnything combines the LoRA weight updates of individual target concepts through normalized summation ΔWEA=1N∑i=1NΔWi\Delta \mathbf{W}_{EA}=\frac{1}{N} \sum _{i=1}^{N} \Delta \mathbf{W}_i. Therefore, our implementation follows its original FLUX-based setting without requiring architectural adaptation.

MACE, on the other hand, was originally designed for U-Net-based diffusion models. However, our comparison does not aim to reproduce its entire concept-erasure training framework on FLUX. Instead, we focus on its LoRA merging strategy for compositional multi-concept erasure. Specifically, we apply the adapted MACE merging strategy to the same independently trained single-concept erasure LoRAs used by Mosaic, rather than training separate LoRAs using MACE.

Therefore, all methods use the same underlying single-concept erasers, allowing us to fairly compare their compositional erasure strategies without introducing differences in the quality of the individual erasers. This controlled comparison is intended to evaluate the effectiveness of different composition strategies rather than the original end-to-end concept-erasure training frameworks. Please note that Section A provides the aforementioned details.

W.3) Missing comparisons with simpler blending strategies

We thank the reviewer for this valuable suggestion. Following the reviewer's suggestion, we implement two additional baselines and compare them with Mosaic. Let vbasev_{\mathrm{base}} denote the vector field predicted by the original FLUX model, and viv_i denote the vector field predicted by the model with the (i)-th concept-specific erasure LoRA.

  • Direct Vector Field Averaging: We directly average the vector fields predicted by independently trained concept-specific erasers, without using spatial masks or discrepancy-based weighting: vavg(p)=1N∑i=1Nvi(p),v_{\mathrm{avg}}(p)=\frac{1}{N}\sum_{i=1}^{N}v_i(p), where NN is the number of target concepts and pp denotes a spatial location. This baseline assigns equal weights to all concept-specific vector fields across all spatial locations.

  • Blending without Spatial Masks: We remove the spatial masks from Mosaic while retaining discrepancy-based weighting. Specifically, we first compute the vector-field discrepancy for each concept as di(p)=∥vi(p)−vbase(p)∥2d_i(p)=\lVert v_i(p)-v_{\mathrm{base}}(p)\rVert_2. We then normalize the discrepancies across all target concepts as ri(p)=di(p)∑j=1Ndj(p)+ϵr_i(p)=\frac{d_i(p)}{\sum _{j=1}^{N}d_j(p)+\epsilon}.

    We then blend the vector fields without applying spatial masks, using vNM(p)=vbase(p)+∑i=1Nri(p)vi(p)1+∑i=1Nri(p)v_{\mathrm{NM}}(p)=\frac{v_{\mathrm{base}}(p)+\sum _{i=1}^{N}r_i(p)v_i(p)}{1+\sum _{i=1}^{N}r_i(p)}.

    Unlike direct averaging, this baseline assigns spatially varying weights based on the relative discrepancies, but does not explicitly restrict the intervention regions using spatial masks.

We conduct these experiments using the same 252 prompts described in W.1, following the same evaluation protocol as our main experiments. We evaluate the Erasure Success Rate (ESR) using Gemma4, Selective Alignment (SA), Aesthetic score.

MethodESR-Gemma(↑)SA(↑)Aesthetic(↑)
Direct Vector Field Averaging49.2195.985.97
Blending without Spatial Masks10.7195.535.95
Mosaic (Ours)79.495.985.91

As shown in the table, Mosaic achieves the highest ESR (79.4%), substantially outperforming Direct Vector Field Averaging (49.21%) and Blending without Spatial Masks (10.71%). In particular, the significant performance degradation observed when spatial masks are removed highlights the importance of spatially selective intervention. These results demonstrate that discrepancy-based weighting alone is insufficient for effective compositional multi-concept erasure, and that spatial masking plays a crucial role in reducing interference between concept-specific vector fields and improving erasure effectiveness. We will add these results during revision.

W.4) Low performance on challenging multi-concept settings

We thank the reviewer for highlighting this limitation. We acknowledge that Mosaic achieves lower ESR than MACE in certain challenging compositional settings, particularly when the number of target concepts increases. However, as shown in our quantitative results in Tables 8 and 9, MACE exhibits substantially lower Selective Alignment (SA) scores than Mosaic, indicating that MACE tends to interfere more with the generation of non-target concepts. This suggests that the higher ESR achieved by MACE in these settings may come at the cost of preserving unrelated scene content.

To further investigate this trade-off, we conduct additional evaluations of non-target (NT) region preservation in the three-object and one-character + two-object settings using LPIPS, SSIM, and two DINOv2-based similarity metrics: DINO and DINO-patch [1],[2]. Specifically, we use SAM3 to identify non-target regions and obtain corresponding spatial masks from images generated by the original FLUX model. We then apply the same masks to the original FLUX-generated images and the corresponding erased images to assess how faithfully each method preserves the original appearance and structure of non-target content. DINO measures the cosine similarity between features averaged over non-target patches, while DINO-patch averages the cosine similarities between spatially corresponding non-target patch features. Both DINOv2-based metrics use the same non-target mask derived from the original image.

MethodNT-LPIPS (↓)NT-SSIM (↑)NT-DINO (↑)NT- DINO patch (↑)
MACE0.60310.46470.70700.3670
Mosaic (Ours)0.57720.46610.73560.3953

These results indicate that Mosaic better preserves non-target regions than MACE, revealing a trade-off between erasure effectiveness and non-target preservation in challenging compositional settings. Nevertheless, the absolute performance of both methods remains limited on these difficult tasks. We therefore agree that improving erasure effectiveness in challenging cases while preserving non-target content is an important direction for future research.

[1] Caron et al., Emerging Properties in Self-Supervised Vision Transformers, ICCV 2021 [2] Ruiz et al., DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation, CVPR 2023

W.5) Limitations of the proposed mask assignment strategy

We thank the reviewer for pointing out this. To investigate how the threshold affects erasure performance, we evaluate τ∈{0.3,0.4,0.5}\tau \in \lbrace 0.3, 0.4, 0.5 \rbrace on both two-concept and three-concept compositions. We select the top four character and object concepts based on their single-concept LoRA erasure performance and constructed compositional prompts from their combinations. For each setting, we evaluate 120 prompts using identical generation configurations across thresholds.

Two-concepts :

τ\tauESR-Qwen (↑)SA (↑)Aesthetic (↑)
0.581.991.95.86
0.483.991.05.84
0.386.390.95.82

Three-concepts :

τ\tauESR-Qwen (↑)SA (↑)Aesthetic (↑)
0.525.289.16.05
0.436.987.56.08
0.341.187.35.99

In both settings, lowering the threshold improves ESR but slightly reduces SA, indicating a trade-off between erasure effectiveness and non-target preservation. This trend is particularly pronounced in three-concept compositions, where reducing the threshold from 0.5 to 0.3 increases ESR from 25.2% to 41.1%, while SA decreases from 89.1% to 87.3%. These findings are consistent with the mask assignment analysis, suggesting that a lower threshold can enable more effective erasure when multiple concepts contribute comparable vector-field discrepancies.

We originally selected τ=0.5\tau=0.5 as the default threshold to balance erasure effectiveness and non-target preservation. However, these additional results indicate that a lower threshold has a possibility to improve erasure performance, particularly in more challenging three-concept settings. Therefore, we acknowledge that a fixed threshold may not be optimal across all compositional settings. Developing an adaptive threshold selection strategy that accounts for the number and characteristics of target concepts is a promising direction for future work. We will add the aforementioned analysis in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.