acceptodds
Under review as a conference paper at ICLR 2027

Beyond Named Prompts: Mitigating Direct and Indirect IP Generation via Generation-State Intervention

Abstract

Modern text-to-image diffusion models can generate images that closely resemble recognizable intellectual-property (IP) characters, whether prompted directly by character names or indirectly through name-free semantic cues. Visual descriptions of appearance, clothing, and composition can induce the latter, challenging mitigations that rely only on prompt inspection or predefined parameter-level interventions. In this work, we introduce a generation-state intervention framework that monitors target-like semantics as they emerge during diffusion and redirects the same generation trajectory without modifying model parameters. The framework uses a predicted clean output for early monitoring and a generic-category-anchored contrast to reduce target-related components while anchoring the requested visual content. Across FLUX.1-dev and Stable Diffusion 3.5 Medium, our method substantially reduces target-character detections under both direct and indirect prompting while maintaining consistency with the visual characteristics specified by the prompt. Together, these results show that the evolving generation state provides an actionable signal for mitigation even when the prompt contains no explicit target reference. Our code is available at https://anonymous.4open.science/r/DiT-IP-Generation-Mitigation-8C43.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Transformer, Text-to-Image Models, Copyright Infringement

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_5gFKOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the suppression of specific IP characters in text-to-image models, focusing on cases where prompts induce character generation through appearance descriptions without explicitly naming the character. The method uses the model’s velocity prediction to estimate the predicted clean output, and monitors the emerging character identity through similarity to a reference images. When the similarity exceeds a threshold, the method computes the difference between predictions conditioned on the target character and its generic category at the current generation state, then subtracts the positive projection along this difference direction from the original prediction. The generic-category anchor aims to suppress character identity while preserving the subject category.

The authors evaluate the method on FLUX.1-dev and SD 3.5 Medium under both direct and indirect prompting. Results show lower mean target-character detection rates than the compared baselines across all four settings, while generic-characteristic consistency remains close to that of the original models. Additional composition experiments suggest that the method can complement mitigation strategies applied at other stages, although it substantially increases inference cost.

Critical strengths

S1. The problem is clearly formulated and practically relevant to indirect prompting in settings where prompts cannot be modified. Define risk in terms of the generated output and its evolving state, rather than restricting it to whether the prompt contains a character name is well motivated.

S2. The generic-category anchor is well motivated and supported by an ablation study. Comparing the null-condition anchor with the generic-category anchor provides stronger evidence than reporting the effectiveness of the full method alone and supports the rationale that preserving category-level semantics helps reduce the side effects of intervention.

S3. The paper is relatively comprehensive, with supplementary analyses of composability, computational cost, and evaluator reliability. These analyses strengthen the paper and help readers assess the method’s practical limitations.

Critical weaknesses

W1. The method assumes that the target character is known in advance: for each generation, the system already knows that “Olaf is the character to suppress.” However, a key challenge of indirect prompting is that the system does not know which character the user is attempting to elicit. In reality, is the method provided with a single target character in advance for each test case? Under indirect prompting, is this target supplied by the evaluation labels or identified automatically by the system? Extending the method from a single specified character to a set of protected characters, without knowing which character each query targets, would substantially improve its practical utility.

W2. Indirect prompts consist of a character’s distinctive attributes, such as “a carrot nose and twig arms.” Yet these are precisely the attributes encoded by d=vOlaf−vsnowmand = v_{\text{Olaf}} - v_{\text{snowman}}. Removing the component along this direction necessarily weakens content explicitly requested by the user. The paper does not define what an “ideal output” should look like under indirect prompting: which attributes must be preserved to remain faithful to the prompt, and which can be removed to eliminate character identity. The method may therefore reduce the recognizability of protected characters by sacrificing some of the user’s requested content.

W3. The default intervention strength introduces reverse guidance beyond orthogonal projection removal. When ⟨q,d⟩>0\langle q,d\rangle>0, the corrected component along the target direction is (1−λ)ad(1-\lambda) ad. Thus, the default 1.5 not only removes the positive projection but also introduces a component in the opposite direction. This operation is consistent with the optimization constraint in Eq. 7, but should be clearly distinguished from exact projection removal (λ=1\lambda=1). Although Table 6 reports an ablation on intervention strength, the paper could further discuss how this additional reverse guidance affects specific attributes requested in the prompt and assess the extent to which stronger suppression comes at the cost of greater semantic drift.

Questions for authors

Please address the points that directly correspond to the concerns raised in the Weaknesses section. No new experiments are required.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5gFK (1)

We sincerely thank the reviewer for the time and care devoted to reviewing our work. We appreciate the reviewer's recognition of the generation-state-based formulation of risk for indirect prompts, the ablation evidence for generic-category anchors, and the comprehensive experimental analyses. Below, we address the reviewer's comments individually.

W1: Target IP known in advance?

Thank you for the insightful comment. Yes. In our main experiments, following prior works [1, 2], each test case is assigned a single target IP by the evaluation setup, including under Indirect prompting; the system does not automatically identify the target in these experiments. We agree that removing this per-query target prior improves practical utility, and have therefore implemented and evaluated a small-scale multi-IP extension. This extension largely retains single-target suppression and CONS with little additional inference cost, without requiring per-query target specification.

In the multi-IP setting, the deploying platform or rights holder specifies the protected set in advance, and we prepare a reference bank and generic-category anchor for each IP offline. The system does not need to know which protected IP a query targets. Instead, the monitor compares the predicted clean output with all reference banks and applies the projection for the highest-scoring IP above threshold.

To validate this capability, we compare single-target and multi-IP mitigation on FLUX.1 with nested sets of 3, 5, and 10 randomly selected protected IPs, using matched prompts and three seeds for each IP (Table 11, Appendix B.10). Arrows denote single-target to multi-IP results, and DETECT is reported as detected images / total images. Multi-IP mitigation maintains CONS close to single-target values, with unchanged DETECT except for the 10-IP Indirect case (5/30 to 7/30).

Protected IPsDirect DETECT ↓Direct CONS ↑Indirect DETECT ↓Indirect CONS ↑
30/9 → 0/90.92 → 0.920/9 → 0/90.86 → 0.86
50/15 → 0/150.81 → 0.811/15 → 1/150.81 → 0.80
100/30 → 0/300.80 → 0.805/30 → 7/300.81 → 0.81

We also report inference time and the fraction of applied interventions directed at another protected IP (non-target updates) in Table 12. Multi-IP mitigation adds only 0.10–0.41 s/img. Although some updates are directed at other protected IPs, suppression and CONS remain close to single-target results in these tests.

Protected IPsDirect time (s/img)Non-target updatesIndirect time (s/img)Non-target updates
313.10 → 13.411.5%13.24 → 13.3925.7%
512.26 → 12.622.4%13.13 → 13.2215.9%
1012.12 → 12.406.3%12.86 → 13.2713.3%

[1] He, L., et al. "Fantastic Copyrighted Beasts and How (Not) to Generate Them." ICLR, 2025.

[2] Han, N., et al. "GrOCE: Graph-Guided Online Concept Erasure for Text-to-Image Diffusion Models." CVPR, 2026.

W2: What an ideal output should look like.

Thank you for raising another insightful question. Yes, the target IP versus generic category contrast can include semantics that overlap with explicitly requested appearance attributes. The ideal output is a generic subject that retains the requested content without being recognizable as the target IP. Given the difficulty existing methods face in suppressing Indirect IP generation, our design prioritizes suppression when it conflicts with prompt adherence, while using the generic category as a preservation reference and retaining as much requested content as possible.

We would also like to clarify that using a generic-category anchor does not mean reducing every output to generic-category content and erasing all user-specified attributes. The original user prompt continues to condition denoising, allowing requested details beyond the generic category to remain. We do not prescribe that attributes such as a carrot nose or twig arms must be removed. In Fig. 6, the snowman generated from the Olaf-related prompt retains both, illustrating the desired balance between suppressing target identity and preserving requested attributes. Fig. 10 similarly shows a generic fish retaining the requested line-art rendering, and the semantic and human evaluations provide further evidence of content retention (Tables 15–16, Appendix B.11).

Nevertheless, the boundary between shared attributes and recognizable target identity can be subtle: faithfully reproducing every requested attribute may itself recreate the protected IP. Fine-grained control over this boundary remains an open challenge. We hope this work motivates more precise mitigation strategies that preserve more indirect attributes without recreating the target IP, allowing service providers to choose the degree of suppression according to their platform policies.

AuthorsOpenReview ↗

Response to Reviewer 5gFK (2)

W3: Projection strength 1.5 introduces reverse guidance.

Yes. The reverse component is intentional: it provides stronger suppression than removing the aligned component alone. To limit inference overhead and unnecessary semantic changes, we intervene only at selected steps, while the original user prompt continues to drive generation toward the target throughout denoising. A moderate reverse correction at these selected steps therefore helps counter the continuing generation bias.

As you note, the strength ablation supports λ=1.5\lambda=1.5 as a balance between suppression and characteristic consistency (Fig. 7; Table 8, Appendix B.5). It improves suppression over λ=1\lambda=1 across both models and prompt types with only small CONS decreases, whereas increasing to 2.0 offers less consistent benefits. We have also revised the intervention-strength discussion in the manuscript (Sec. 4.2) to clarify the rationale for the reverse correction and our choice of λ=1.5\lambda=1.5.

I thank the authors for their detailed response, which has clarified all my questions. Although the issue I raised remains, I find the paper sufficiently persuasive given its stated research objectives. I will therefore keep my original rating of 4.

Thank you for your positive feedback. We appreciate your recognition of the paper’s contribution. We also appreciate your constructive comments, which have helped us clarify the scope and limitations of our work.

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

This paper proposes a training-free method to mitigate recognizable IP-character generation in flow-matching text-to-image diffusion models, including cases prompted indirectly without character names. During sampling, it decodes predicted clean outputs and uses DINOv2 similarity to detect emerging target identities. When a threshold is exceeded, it suppresses target-aligned predictions using a generic-category anchor. Experiments on FLUX.1-dev and SD 3.5 Medium using CopyCat show lower target detection rates than four mitigation baselines while largely preserving characteristic consistency. The paper also reports ablations, mitigation composition, efficiency, generation quality, and evaluator validation.

Critical strengths
  1. The paper addresses a practically relevant gap with a simple, training-free mechanism. The paper clearly identifies the limitations of prompt-level prevention and post-generation detection. It demonstrates that the predicted clean output can reveal target identity early in sampling, enabling intervention within the same generation trajectory. The method requires only 0.42 seconds of per-target preparation and no parameter updates. Furthermore, combining it with existing mitigations consistently reduces DETECT. These results suggest that generation-state intervention can serve as a complementary safeguard alongside existing approaches.

  2. The generic-category anchor is a well-motivated and well-ablated design choice. Using a null anchor can reduce target detection but also damage the requested subject, while the generic-category anchor recovers characteristic consistency. This is one of the clearest methodological insights of the paper. It demonstrates the importance of selectively suppressing character identity rather than removing all target-related semantics.

Critical weaknesses

1. Insufficient evidence to establish the methodological novelty over existing inference-time guidance methods. The paper proposes a monitoring-triggered correction that removes the target-aligned component from the model prediction using a generic-category anchor. However, this correction is closely related to inference-time negative guidance, with an adaptive projection coefficient. The evaluated baselines primarily operate at other stages of generation. Although TRIM uses negative prompting during regeneration, it does not perform guidance within the same generation trajectory. Moreover, the paper discusses existing inference-time guidance methods, including Guardians of Generation [3], which also uses adaptive guidance for copyright protection, but does not compare against them. As a result, it remains unclear whether the reported improvements come from the proposed monitoring and projection design or from applying negative guidance during sampling. This limits the evidence for the method's novelty and is my main concern with the paper.

2. Insufficient evaluation of image quality and prompt fidelity after intervention. The paper evaluates content preservation using CONS and reports CLIP Score, VQAScore, and FID on COCO-10k. However, the COCO experiment triggers the intervention on only 5 of 10,000 FLUX.1 images and none of the SD 3.5 images. Therefore, these results mainly demonstrate a low false-trigger rate rather than the quality of images that actually undergo intervention. Meanwhile, CONS evaluates only a short generic characteristic instead of the full prompt. It cannot adequately capture changes in artistic style, scene details, or other requested attributes. For example, Fig. 11 shows that Winnie the Pooh changes from a cartoon character to a more realistic bear. Since the method aims to suppress character identity while preserving other requested content, the lack of direct quality evaluation on intervened images limits the strength of its preservation claims.

3. Insufficient evaluation of the monitoring mechanism's robustness across different visual conditions. The monitor compares global DINOv2 features against four reference images with a fixed threshold, while the final DETECT metric relies on a VLM-based evaluator. For small characters, cluttered scenes, or stylized depictions (e.g., oil paintings, line art, or plush toys), the similarity may remain below the threshold even when the final evaluator recognizes the target. This potential mismatch between the monitoring and evaluation criteria raises concerns about the reliability of the intervention across different visual conditions.

Questions for authors
  1. Could the authors compare their method against existing inference-time guidance approaches, particularly Guardians of Generation [3], SLD [1], and DNG [2], where applicable?

  2. Could the authors evaluate image quality specifically on samples where the intervention is activated? For example:

    • CLIP Score or VQAScore against descriptions of the requested content with identity-specific references removed.
    • Human evaluation of visual quality, artistic style, and preservation of non-IP-related prompt details.
  3. The current experiments monitor each prompt only against its ground-truth target character. However, a practical deployment would need to monitor multiple protected characters without knowing which one might appear. Could the authors provide a small-scale experiment with simultaneous monitoring against multiple target reference banks, reporting DETECT, CONS, false-trigger rate, and inference overhead?

  4. Could the authors evaluate monitoring robustness for small characters, cluttered scenes, and stylized depictions, reporting both monitoring trigger rates and final DETECT? This would help clarify whether the monitoring mechanism reliably identifies the characters recognized by the final evaluator.

If the authors can address my main concerns, I would be willing to raise my score.

References

[1] Schramowski, P., et al. "Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models." CVPR, 2023.

[2] Koulischer, F., et al. "Dynamic Negative Guidance of Diffusion Models." ICLR, 2025.

[3] Roy, S., et al. "Guardians of Generation: Dynamic Inference-Time Copyright Shielding with Adaptive Guidance for AI Image Generation." ICCV Workshops, 2025.

4 comments
AuthorsOpenReview ↗

Response to Reviewer bDau (1)

We sincerely thank the reviewer for the careful review and constructive feedback. We appreciate the reviewer's recognition of the practical relevance of our training-free approach, its lightweight target-specific preparation and compatibility with existing defenses, and the motivation and ablation evidence for generic-category anchors. Below, we address the reviewer's comments individually.

W1 / Q1: Novelty over existing inference-time guidance.

Our novelty lies in combining (1) visual-reference monitoring to determine when to intervene with (2) generic-category-anchored projection to determine where to steer generation, beyond GoG's prompt-embedding modification and SLD/DNG's target-only negative guidance. We have added comparisons with all three: Ours achieves the lowest Indirect DETECT on both models, with higher CONS than GoG and comparable CONS to SLD/DNG. GoG achieves lower Direct DETECT, but with lower CONS (0.55/0.63 vs. our 0.75/0.74).

For these three methods, we evaluate IP suppression (DETECT) and characteristic consistency (CONS) under the main evaluation settings (Table 10, Appendix B.9). Specifically, GoG [1] detects protected concepts in the prompt, rewrites it, and mixes the original and rewritten embeddings before sampling. Therefore, its mitigation stage is more like GrOCE instead of TRIM. Our original Related Work did not distinguish this prompt-embedding intervention from sampling-time guidance. We have corrected its placement in Sec. 5.

We also adapt SLD [2] and DNG [3], two inference-time guidance methods, using the target IP name as the suppressive concept. SLD applies its negative guidance without visual detection, while DNG adjusts guidance strength through a time- and state-dependent posterior estimate on model latent outputs. Neither uses visual-reference feedback or a category-relative correction. On Direct prompts, identical user prompt and negative guidance conditions (both IP names) are used, nullifying SLD's suppressive difference. DNG also shows limited suppression in this setting despite its adaptive strength.

ModelMethodDirect DETECT ↓ (%)Direct CONS ↑Indirect DETECT ↓ (%)Indirect CONS ↑
FLUX.1Vanilla98.5 ± 1.30.74 ± 0.0170.4 ± 2.60.77 ± 0.01
SLD100.0 ± 0.00.71 ± 0.0268.1 ± 2.60.74 ± 0.00
DNG97.0 ± 2.60.73 ± 0.0166.7 ± 3.80.77 ± 0.00
GoG8.9 ± 3.80.55 ± 0.0137.8 ± 2.20.66 ± 0.01
Ours16.3 ± 3.40.75 ± 0.0230.4 ± 3.40.76 ± 0.01
SD 3.5Vanilla90.2 ± 7.30.72 ± 0.0259.1 ± 2.30.76 ± 0.02
SLD90.9 ± 2.30.72 ± 0.0237.1 ± 6.60.77 ± 0.00
DNG84.8 ± 1.30.72 ± 0.0157.6 ± 6.90.76 ± 0.00
GoG11.4 ± 2.30.63 ± 0.0223.5 ± 5.20.71 ± 0.03
Ours15.2 ± 5.70.74 ± 0.0315.2 ± 3.50.75 ± 0.01

To further isolate the role of monitoring, we also compare our triggered correction with applying the same correction at every eligible step on FLUX.1 (Table 6, Appendix B.3). Every-step intervention lowers Direct/Indirect DETECT further (1.5%/9.6% vs. 16.3%/30.4%), but also lowers CONS (0.70/0.74 vs. 0.75/0.76) and increases inference time (14.19 vs. 13.36 s/img). This supports monitoring as a means of balancing suppression, characteristic consistency, and inference cost.

Together, these comparisons and the monitoring ablation support our design beyond simply applying negative guidance during sampling. Visual-reference feedback provides an intervention signal from emerging target appearance, particularly for Indirect prompts that omit the IP name, while the generic-category anchor provides a reference for retaining general subject semantics.

[1] Roy, S., et al. "Guardians of Generation: Dynamic Inference-Time Copyright Shielding with Adaptive Guidance for AI Image Generation." ICCV Workshops, 2025.

[2] Schramowski, P., et al. "Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models." CVPR, 2023.

[3] Koulischer, F., et al. "Dynamic Negative Guidance of Diffusion Models." ICLR, 2025.

AuthorsOpenReview ↗

Response to Reviewer bDau (2)

W3 / Q4: Monitoring robustness in challenging conditions.

To test robustness beyond the original Indirect prompts, we compare nine methods on FLUX.1 under variations in target scale (Local Region), rendering style/material (Oil Painting, Line Art, and Plush Toy), and scene complexity (Complex Scene). We combine Indirect descriptions with the corresponding condition instructions for 10 sampled IPs and three seeds, yielding 30 images per condition per method. Local Region requests a small figure occupying 5–10% of the image on a simple background; Complex Scene places a fully visible subject amid clutter. Ours reduces target detection across all five conditions while maintaining high characteristic consistency. The table reports condition-specific DETECT and aggregate DETECT/CONS as mean ± sample standard deviation across seeds (Table 13, Appendix B.11).

MethodLocal RegionOil PaintingLine ArtPlush ToyComplex SceneDETECT (%) ↓CONS ↑
Vanilla70.0 ± 10.083.3 ± 5.873.3 ± 5.883.3 ± 5.863.3 ± 5.874.7 ± 1.20.76 ± 0.02
CopyCat Rewrite60.0 ± 0.066.7 ± 11.533.3 ± 5.866.7 ± 5.853.3 ± 11.556.0 ± 4.00.81 ± 0.01
GrOCE23.3 ± 5.833.3 ± 5.816.7 ± 5.820.0 ± 10.00.0 ± 0.018.7 ± 2.30.70 ± 0.00
TRIM36.7 ± 5.856.7 ± 5.830.0 ± 10.050.0 ± 0.050.0 ± 0.044.7 ± 2.30.76 ± 0.02
EraseAnything43.3 ± 11.546.7 ± 5.850.0 ± 10.063.3 ± 5.860.0 ± 10.052.7 ± 1.20.71 ± 0.01
SLD0.0 ± 0.063.3 ± 5.826.7 ± 20.876.7 ± 15.356.7 ± 11.544.7 ± 4.20.60 ± 0.02
DNG73.3 ± 5.873.3 ± 5.866.7 ± 5.870.0 ± 10.063.3 ± 15.369.3 ± 1.20.75 ± 0.00
GoG0.0 ± 0.090.0 ± 17.310.0 ± 10.060.0 ± 0.026.7 ± 5.837.3 ± 4.60.52 ± 0.01
Ours33.3 ± 5.823.3 ± 11.526.7 ± 15.343.3 ± 5.830.0 ± 10.031.3 ± 5.00.79 ± 0.01

Ours reduces aggregate DETECT from 74.7% to 31.3%, with CONS of 0.79 vs. Vanilla's 0.76. It achieves the lowest Oil Painting DETECT, but other methods suppress more strongly in some conditions. GrOCE has lower aggregate DETECT (18.7%) but also lower CONS (0.70). SLD and GoG reach zero Local Region DETECT, although their aggregate CONS is lower (0.60 and 0.52). In the aggregate DETECT–CONS plot (Fig. 9, panel a), Ours lies on the Pareto frontier among the nine compared methods: none achieves both lower DETECT and higher CONS. Fig. 10 provides qualitative comparisons across all five conditions. For example, in Local Region, Ours retains a generic snowman, whereas GrOCE substitutes an unrelated animal and SLD/GoG remove the subject.

We also report image-level monitor trigger rates (Table 14, Appendix B.11): an image counts as triggered if the monitor crosses the threshold at least once during generation. The following rates use the same three seeds and 30 images per condition, reported as mean ± sample standard deviation (%). The monitor triggers on 83.3% of Local Region images, 86.7% of Plush Toy and Complex Scene images, and all Oil Painting and Line Art images. These results show that visual monitoring continues to trigger when the target is small, stylized, or embedded in a complex scene.

MethodLocal RegionOil PaintingLine ArtPlush ToyComplex Scene
Ours83.3 ± 5.8100.0 ± 0.0100.0 ± 0.086.7 ± 5.886.7 ± 5.8

These additional experiments demonstrate the applicability of our visual monitoring and category-anchored intervention beyond the original Indirect prompts. We believe that developing dedicated defenses for complex visual conditions is a valuable extension of future work.

AuthorsOpenReview ↗

Response to Reviewer bDau (3)

W2 / Q2: Image quality and prompt fidelity.

Thank you for mentioning this point. However, the original Direct prompts name the target IP, while Indirect prompts primarily describe its identifying appearance. Scoring against these texts would conflate target-identity retention with prompt fidelity, and they provide little independent style or scene content to assess. We therefore add complex-scene VQAScore evaluation and five-rater human evaluations of quality, requested style, and scene fulfillment on images from the challenging-condition experiments. Ours substantially reduces target detection while retaining VQAScore and human ratings close to Vanilla.

Automated evaluation of semantic preservation

On the Complex Scene subset of the test above (10 IPs ×\times 3 seeds per method), we compare nine methods using VQAScore against target-name-free descriptions of the generic category subject and requested scene. This metric assesses subject and scene fulfillment alongside DETECT and CONS (Table 15, Appendix B.11).

MethodDETECT (%) ↓CONS ↑VQAScore ↑
Vanilla63.3 ± 5.80.74 ± 0.040.75 ± 0.04
CopyCat Rewrite53.3 ± 11.50.77 ± 0.020.76 ± 0.04
GrOCE0.0 ± 0.00.62 ± 0.020.63 ± 0.03
TRIM50.0 ± 0.00.81 ± 0.020.83 ± 0.04
EraseAnything60.0 ± 10.00.71 ± 0.020.70 ± 0.02
SLD56.7 ± 11.50.77 ± 0.000.76 ± 0.03
DNG63.3 ± 15.30.76 ± 0.010.70 ± 0.02
GoG26.7 ± 5.80.70 ± 0.040.67 ± 0.03
Ours30.0 ± 10.00.76 ± 0.030.74 ± 0.05

Ours reduces DETECT from 63.3% to 30.0% while retaining VQAScore close to Vanilla (0.74 vs. 0.75), placing it on the DETECT–CONS and DETECT–VQAScore Pareto frontiers (Fig. 9, panels b,c). TRIM has higher CONS and VQAScore but leaves more detections (50.0%), whereas GrOCE reaches 0.0% DETECT with lower CONS and VQAScore. Fig. 10 illustrates how mitigation can sacrifice the intended content: GrOCE replaces Olaf with an unrelated animal, TRIM/GoG remove the line-art subject, and SLD blurs it. Ours retains a generic snowman and a fish in the requested line-art style in these examples.

Human evaluation of quality, style, and non-IP-related prompt details

To assess perceptual quality and requested content directly, we sample two images each from Oil Painting, Line Art, and Plush Toy, plus six from Complex Scene per method, using matched requests across the nine methods (108 images). Five human raters assess overall quality on all sampled images, style adherence on the six stylized images, and scene details on the six complex-scene images per method. Quality and style use 1–5 scales, while scene fulfillment measures the presence of four requested objects: books, a ceramic mug, a potted plant, and a desk lamp (Table 16, Appendix B.11).

MethodQuality ↑Style ↑Scene fulfillment (%) ↑DETECT (%) ↓CONS ↑
Vanilla4.724.2085.074.70.76
CopyCat Rewrite4.784.3090.056.00.81
GrOCE4.784.2090.018.70.70
TRIM4.154.1783.344.70.76
EraseAnything4.684.2785.052.70.71
SLD3.673.3093.344.70.60
DNG4.803.7073.369.30.75
GoG3.223.2080.837.30.52
Ours4.654.2385.031.30.79

Ours scores close to Vanilla in quality (4.65 vs. 4.72), requested style (4.23 vs. 4.20), and scene fulfillment (both 85.0%), while DETECT falls from 74.7% to 31.3%. Compared with TRIM, GoG, and SLD, Ours has higher quality and style ratings alongside lower DETECT. CopyCat Rewrite scores higher on all three human dimensions but leaves more target detections (56.0%). GrOCE achieves lower DETECT and higher quality and scene scores, while Ours retains higher CONS (0.79 vs. 0.70) and similar style ratings.

Together, the automatic and human evaluations show that Ours substantially reduces target identity while largely preserving the generic subject, requested style, and scene content.

AuthorsOpenReview ↗

Response to Reviewer bDau (4)

Q3: Multi-IP mitigation.

Thank you for your insightful comment. Our main experiments follow the target-specific evaluation setting used in prior work [1, 2]. Our method also supports multi-IP mitigation without knowing the query's target in advance, with comparable suppression and CONS at little additional inference cost in our tests.

Given a reference bank and generic-category anchor for each protected IP, we monitor all banks and intervene using the highest-scoring IP above threshold. The query's target label is used only for evaluation. We compare single-target and multi-IP mitigation on FLUX.1 with nested sets of 3, 5, and 10 randomly selected protected IPs, using matched prompts and three seeds for each IP. Table 11 (Appendix B.10) reports the results. Arrows denote single-target → multi-IP results; DETECT is reported as detected images / total images. Multi-IP mitigation maintains CONS close to single-target values, with unchanged DETECT except for the 10-IP Indirect case (5/30 → 7/30).

Protected IPsDirect DETECT ↓Direct CONS ↑Indirect DETECT ↓Indirect CONS ↑
30/9 → 0/90.92 → 0.920/9 → 0/90.86 → 0.86
50/15 → 0/150.81 → 0.811/15 → 1/150.81 → 0.80
100/30 → 0/300.80 → 0.805/30 → 7/300.81 → 0.81

We also measure inference time and the fraction of interventions directed at an IP other than the intended target (non-target updates). Table 12 reports both measures. Multi-IP mitigation adds only 0.10–0.41 s/img. Although some interventions are directed at other protected IPs, the results show little change in suppression and characteristic consistency relative to single-target mitigation.

Protected IPsDirect time (s/img)Non-target updatesIndirect time (s/img)Non-target updates
313.10 → 13.411.5%13.24 → 13.3925.7%
512.26 → 12.622.4%13.13 → 13.2215.9%
1012.12 → 12.406.3%12.86 → 13.2713.3%

[1] He, L., et al. "Fantastic Copyrighted Beasts and How (Not) to Generate Them." ICLR, 2025.

[2] Han, N., et al. "GrOCE: Graph-Guided Online Concept Erasure for Text-to-Image Diffusion Models." CVPR, 2026.

Reviewer_2ebxOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an inference-time method to reduce recognizable IP character generation from direct and indirect prompts in text-to-image diffusion models. The method monitors the emerging image during the generation process to detect whether it resembles a target IP character. When the similarity exceeds a threshold, it adjusts the generation process to reduce character-specific identity while preserving general visual characteristics. Experiments on two diffusion models show substantial reductions in character detection rates while maintaining similar generic-characteristic consistency.

Critical strengths
  1. The proposed method is well motivated. The paper gives a clear discussion of existing IP mitigation methods and their limitations, especially the timing dilemma between early prevention and late detection.
  2. The paper provides strong experimental support for its approach. The observations related to emerging IP character identity during image generation are well described and supported by Figure 3, while comparisons with existing methods, ablation studies, and evaluations of image quality and computational efficiency demonstrate the effectiveness of the proposed method across two diffusion models.
  3. The paper combines predicted-output monitoring with anchor-conditioned intervention, which is an interesting contribution that reduces IP-specific identity while preserving general visual characteristics during image generation.
Critical weaknesses
  1. The paper does not clearly demonstrate the benefit of the adaptive intervention mechanism. Although Figures 3 and 9 show observations supporting adaptive monitoring, the authors should provide a clearer justification for their choice of threshold.
  2. The paper discusses “Guardians of Generation” by Roy et al. (2025) in the related work section but does not directly compare it with the proposed method, although both use intervention during image generation, leaving their relative effectiveness unclear. The authors should clarify the methodological differences and, if feasible, compare both methods experimentally.
  3. Figure 5 presents the main framework, but the connections between stages and the initial input are unclear. Furthermore, the terminology used in the panel titles and figure caption is inconsistent. The authors could clarify the information flow and use consistent terminology throughout the figure.
Questions for authors
  1. The paper uses a fixed monitoring threshold of 0.5 without clearly explaining its selection. Could the authors justify this choice and compare different thresholds, as well as intervention at every eligible step, in terms of DETECT, CONS, and inference time?
  2. The paper states that the VLM receives only the target IP name to generate the generic-category anchor. Could the authors clarify why a VLM is used instead of a text-only LLM?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 2ebx (1)

We sincerely thank the reviewer for the time and effort devoted to reviewing our paper. We appreciate the reviewer's recognition of our motivation, the value of combining predicted-output monitoring with generic-category-anchored intervention, and the experimental support across FLUX.1 and SD 3.5. Below, we address the reviewer's comments.

W1 / Q1: Monitoring threshold and adaptive intervention.

We select .5 to suppress target-IP generation while limiting unnecessary intervention on non-IP content and keeping moderate inference cost. As you note, Fig. 3 suggests a range separating Direct/Indirect generations from Generic Category and COCO controls. We provide two additional FLUX.1 experiments (Appendix B.3).

First, we measure the fraction of Vanilla trajectories crossing each threshold (Table 5). Moving from .6 to .5 increases Indirect coverage from 80.7% to 92.6%. Moving further to .4 adds only 3.7 percentage points but raises Generic Category crossings from 19.3% to 31.1%, supporting .5 as a practical operating point.

ThresholdDirect crossing (%)Indirect crossing (%)Generic Category crossing (%)
.699.3 ± 1.380.7 ± 4.68.1 ± 3.4
.599.3 ± 1.392.6 ± 2.619.3 ± 4.6
.499.3 ± 1.396.3 ± 1.331.1 ± 3.8

Second, we compare our mitigation at each threshold with intervention at every eligible step, measuring IP suppression (DETECT), characteristic consistency (CONS), and inference time (Table 6).

PolicyDirect DETECT ↓ (%)Direct CONS ↑Indirect DETECT ↓ (%)Indirect CONS ↑Inference (s/img)
Vanilla98.5 ± 1.30.74 ± 0.0170.4 ± 2.60.77 ± 0.016.36
τ = .627.4 ± 5.10.76 ± 0.0145.9 ± 6.80.76 ± 0.0111.90
τ = .516.3 ± 3.40.75 ± 0.0230.4 ± 3.40.76 ± 0.0113.36
τ = .411.1 ± 4.40.75 ± 0.0223.7 ± 1.30.76 ± 0.0113.75
Every eligible step1.5 ± 1.30.70 ± 0.019.6 ± 4.60.74 ± 0.0014.19

Lower thresholds encourage more frequent intervention and auxiliary model evaluations, increasing inference cost. Inference time rises from 11.90 s/image at .6 to 13.75 at .4 and 14.19 for every-step intervention. More frequent intervention lowers DETECT but also CONS: from .5 to every-step intervention, Direct/Indirect DETECT falls from 16.3%/30.4% to 1.5%/9.6%, while CONS falls from 0.75/0.76 to 0.70/0.74. This further supports .5 as a practical balance of suppression, selectivity, and cost.

AuthorsOpenReview ↗

Response to Reviewer 2ebx (2)

W2: Comparison with GoG and other adaptable inference-time guidance methods.

We compare with GoG [1] and additionally evaluate two inference-time guidance methods, SLD [2] and DNG [3]. Ours achieves the lowest Indirect DETECT on both FLUX.1 and SD 3.5, with higher CONS than GoG and comparable CONS to SLD/DNG. GoG achieves lower Direct DETECT but also lower CONS.

Our original discussion did not specify GoG's intervention stage. GoG detects protected concepts in the prompt, rewrites it, and mixes the original and rewritten embeddings before sampling. We now group GoG with conditioning-based controls in Related Work (Sec. 5) and explain its prompt-embedding mechanism in Appendix B.9.

To compare with inference-time mitigations, we additionally adapt SLD and DNG to IP suppression. During sampling, SLD subtracts a gated safety-concept (target-IP name) direction from classifier-free guidance, while DNG adjusts negative-guidance strength using a time- and state-dependent posterior estimate.

Our contribution addresses (1) when to intervene through visual-reference feedback from the predicted clean output, which is absent from these baselines, and (2) where to steer generation through generic-category-anchored projection. The latter isolates target-specific deviation relative to the category to retain general subject semantics, beyond embedding mixing or negative guidance.

We evaluate DETECT and CONS under the main evaluation settings (Table 10, Appendix B.9). Compared with GoG, our Indirect DETECT falls from 37.8% to 30.4% on FLUX.1 and from 23.5% to 15.2% on SD 3.5, while CONS rises from 0.66/0.71 to 0.76/0.75. SLD/DNG retain substantially higher Indirect DETECT. On Direct prompts, identical user prompt and negative guidance conditions (both IP names) make the suppressive terms near zero in our SLD/DNG adaptations, leaving their suppresion strength limited.

ModelMethodDirect DETECT ↓ (%)Direct CONS ↑Indirect DETECT ↓ (%)Indirect CONS ↑
FLUX.1Vanilla98.5 ± 1.30.74 ± 0.0170.4 ± 2.60.77 ± 0.01
SLD100.0 ± 0.00.71 ± 0.0268.1 ± 2.60.74 ± 0.00
DNG97.0 ± 2.60.73 ± 0.0166.7 ± 3.80.77 ± 0.00
GoG8.9 ± 3.80.55 ± 0.0137.8 ± 2.20.66 ± 0.01
Ours16.3 ± 3.40.75 ± 0.0230.4 ± 3.40.76 ± 0.01
SD 3.5Vanilla90.2 ± 7.30.72 ± 0.0259.1 ± 2.30.76 ± 0.02
SLD90.9 ± 2.30.72 ± 0.0237.1 ± 6.60.77 ± 0.00
DNG84.8 ± 1.30.72 ± 0.0157.6 ± 6.90.76 ± 0.00
GoG11.4 ± 2.30.63 ± 0.0223.5 ± 5.20.71 ± 0.03
Ours15.2 ± 5.70.74 ± 0.0315.2 ± 3.50.75 ± 0.01

[1] Roy, S., et al. "Guardians of Generation: Dynamic Inference-Time Copyright Shielding with Adaptive Guidance for AI Image Generation." ICCV Workshops, 2025.

[2] Schramowski, P., et al. "Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models." CVPR, 2023.

[3] Koulischer, F., et al. "Dynamic Negative Guidance of Diffusion Models." ICLR, 2025.

W3: Main framework figure.

We have clarified the inputs and closed-loop information flow in Fig. 5 and aligned its panel labels with the caption. Panel (a) supplies the reference bank for monitoring and the generic-category anchor for intervention. Panel (b) shows generation initialized from the prompt and noise, followed by monitoring of the predicted clean output. When triggered, panel (c) applies the projection. The corrected prediction drives the update xt→xt−1x_t\rightarrow x_{t-1}, which feeds back into panel (b), so the next denoising step reflects earlier corrections.

Q2: Generic-category anchor construction with LLM or VLM.

Thank you for the insightful comment. A text-only LLM is viable. We prefer a VLM because its visual priors help identify the target's physical appearance rather than its narrative role, yielding anchors that more accurately describe the generic subject semantics we aim to preserve.

Using the same construction procedure (Appendix A.4), Qwen3-30B-A3B-Instruct-2507 produced 50/50 valid outputs, with 30/50 labels matching our Qwen3-VL-30B-A3B-Instruct anchors. However, for some IPs, like Thor and Mr. Incredible, the VLM returns "human," while the text LLM returns "god" and "superhuman," emphasizing narrative status or abilities. "Human" describes their visible subject category and thus better matches the anchor's purpose.

Thank you for the detailed rebuttal and additional experiments. The new results address most of my concerns, particularly the comparisons with existing guidance methods and the robustness evaluation. I will raise my score accordingly.

I have one remaining clarification question. Were the images used in the new image-quality and human evaluations specifically those where the proposed intervention was actually triggered? Since the method aims to suppress target identity while preserving other requested content, it is important to evaluate image quality specifically after intervention. Including non-triggered samples could overestimate content preservation, as those images remain unaffected by the method. Clarifying this would help determine whether the reported quality results directly support the paper's preservation claims.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overview of Revisions

We sincerely thank the PC, SAC, and AC for their time and effort in overseeing the review process, and all reviewers for their careful reviews and constructive feedback. We appreciate the reviewers' recognition of the practical importance of indirect copyright generation (R#bDau, R#5gFK), the value of predicted-output monitoring (R#2ebx, R#5gFK) and generic-category anchors (R#2ebx, R#bDau, R#5gFK), and the comprehensive experiments and analyses (R#2ebx, R#5gFK).

Below, we summarize the main manuscript revisions and their corresponding reviewer comments, followed by point-by-point response summaries. Changes in the revised manuscript are highlighted in blue for ease of reference. For detailed analysis, please kindly refer to the response box for each reviewer and the corresponding sections of the revised manuscript.

  • Monitoring threshold and adaptive intervention (R#2ebx W1/Q1; R#bDau W1/Q1). Threshold and every-eligible-step comparisons further support our balance of suppression, CONS, and inference cost (Appendix B.3).
  • Comparison with GoG, SLD, and DNG (R#2ebx W2; R#bDau W1/Q1). Comparisons on both models further establish our advantages on Indirect prompts, with method adaptations and GoG's intervention stage clarified (Sec. 5; Appendix B.9).
  • Main framework figure (R#2ebx W3). Fig. 5 now clarifies the inputs and closed-loop information flow.
  • Multi-IP mitigation (R#bDau Q3; R#5gFK W1). Tests with 3, 5, and 10 protected IPs demonstrate mitigation without knowing the query's target in advance (Sec. 4.2; Appendix B.10).
  • Challenging visual conditions (R#bDau W3/Q4). Qualitative comparisons, Pareto plots, and monitor trigger rates further evaluate robustness to target scale, style/material, and scene complexity (Figs. 9-10; Appendix B.11).
  • Image quality and prompt fidelity (R#bDau W2/Q2; R#5gFK W2). Complex-scene VQAScore and five-rater human evaluations provide further evidence of quality, requested style, and scene fulfillment (Tables 15-16, Appendix B.11).
  • Intervention strength (R#5gFK W3). Sec. 4.2 clarifies the rationale for reverse correction and the choice of λ=1.5\lambda=1.5.
AuthorsOpenReview ↗

Summary of Rebuttal Discussion

We thank the reviewers for their constructive feedback and thoughtful follow-up discussion. As the discussion period draws to a close, we briefly summarize their assessments:

  • Reviewer 2ebx (3 → 3): Confirmed that our clarifications, manuscript revisions, and additional experiments satisfactorily addressed their concerns, including threshold selection, comparisons with related methods, and generic-category anchor construction.
  • Reviewer bDau (2 → 3): Stated that the additional experiments addressed most of their concerns and raised their rating. We also provided clarification in response to their follow-up question about evaluating image quality after intervention.
  • Reviewer 5gFK (4 → 4): Confirmed that our response clarified all their questions and found the paper sufficiently persuasive within its stated research objectives, maintaining their clear-acceptance rating.

The revised manuscript incorporates the additional experiments and clarifications summarized in our earlier “Overview of Revisions.” We thank the reviewers again for their time and constructive engagement, which have helped strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.