acceptodds
Under review as a conference paper at ICLR 2027

SPIN: Image Immunization Against Diffusion Editing via Single-Step Projection in Stochastic Neighborhoods

13 views

Abstract

Diffusion models have greatly advanced instruction-guided image editing, while also raising concerns about unauthorized image manipulation. Image immunization addresses this risk by adding imperceptible perturbations to an input image to disrupt subsequent edits. Since editing requests are unknown at image release, protection should remain effective beyond the instruction used to construct the perturbation. Existing immunization methods either require costly full-trajectory backpropagation or use intermediate objectives whose effects may be weakened by subsequent denoising. Meanwhile, a single inference path provides limited feedback about alternative denoising continuations. To address these challenges, we propose SPIN, a framework for image immunization via one-step projection over local stochastic trajectory neighborhoods. Starting from an early denoising state, SPIN generates stochastic neighboring states under the same instruction and predicts their clean latents through one-step projection without full unrolling. We then optimize a bounded input perturbation to maximize the average deviation of these predictions from a clean-edit reference, encouraging the perturbation to disrupt multiple possible editing outcomes. Experiments on two image editors demonstrate substantial gains in protection performance, with SPIN outperforming compared methods across all six metrics under seen instructions and in the more challenging unseen instruction setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Immunization, Diffusion Models, Adversarial Perturbations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JMVpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPIN, an image immunization method that adds small, norm-bounded perturbations to hinder unauthorized diffusion-based editing while preserving the input image’s visual appearance. To reduce the computational cost of full-trajectory backpropagation, SPIN uses single-step projection to predict clean latents from early denoising states. Under a fixed editing instruction, it generates multiple neighboring states using different noise realizations and optimizes the input perturbation to maximize the average cosine distance between their projected clean latents and a fixed clean-edit reference. This reference is obtained by running the full editor once on the unperturbed image. The editing model’s weights remain fixed throughout optimization, and computational overhead is reduced by avoiding full denoising rollouts for each branch. In white-box evaluations targeting SD3 and InstructPix2Pix, SPIN outperforms the evaluated baselines and remains effective under unseen instructions and previously unseen seeds, achieving unseen-instruction immunization success rates of 85% and 91%, respectively.

Critical strengths
  1. By predicting clean latents through single-step projection from early denoising states, SPIN reduces the computational overhead of backpropagating through the entire denoising trajectory at every optimization iteration. This efficiency is particularly compelling relative to methods that differentiate through the full sampling process.

  2. SPIN constructs perturbations using a single instruction by aggregating feedback across multiple stochastic branches, while maintaining strong protection against the tested unseen instructions. Within the same-model white-box setting, its unseen-instruction ISR reaches 85% on SD3 and 91% on InstructPix2Pix, demonstrating effective cross-prompt generalization under the evaluated conditions.

  3. Since protected images are intended for public sharing, perturbation imperceptibility is important. Under the same perturbation budget, SPIN achieves an ISR of 90% on SD3 while maintaining high input fidelity. These results indicate a favorable trade-off between protection effectiveness and input imperceptibility under the evaluated settings.

Critical weaknesses
  1. The evaluation does not establish protection against unseen editing models. Once a protected image is released publicly, the target editing model may be unknown, and robustness to unseen prompts within the same model does not establish cross-model protection. Given that SIFM already evaluates cross-model transfer, experiments applying perturbations constructed on one editor to another without re-optimization would strengthen the practical relevance of SPIN.

  2. The main quantitative comparisons omit relevant methods, including DiffusionGuard, which also targets early denoising stages, and the recent work “Universal Image Immunization against Diffusion-based Image Editing via Semantic Injection.” The latter is particularly relevant to evaluating SPIN’s practical efficiency and scalability because it learns a single universal perturbation reusable across images and evaluates black-box transferability. Given the differences between per-image optimization and universal perturbations, the authors should compare protection performance under matched editing models, model-access conditions, and perturbation budgets, clearly distinguishing initial training costs from per-image application costs, or explain why these methods were excluded.

  3. The adaptation and optimization settings of several baselines are insufficiently specified. For example, SIFM reports excluding MIST and SDS from its SD3 evaluation due to architectural incompatibilities, whereas SPIN reports their SD3 results without explaining how these baselines were implemented or adapted for SD3. To assess the fairness and reproducibility of the comparisons, the authors should clarify whether each baseline was re-optimized against the corresponding target editor and report the model checkpoints, objective adaptations, if any, and method-specific hyperparameters.

Questions for authors
  1. Could the authors evaluate cross-model transfer by applying perturbations generated on one editor to the other editing models without re-optimization?

  2. Could the authors compare SPIN with DiffusionGuard and Lee et al. 2026, under matched evaluation conditions, reporting initial training and per-image application costs separately, or justify their exclusion?

  3. Could the authors clarify how MIST and SDS were implemented or adapted for SD3 and whether each baseline was re-optimized against the corresponding target editor, specifying the checkpoints, any objective modifications, and method-specific hyperparameters?

4 comments

1.Cross-model transferability

Thank you for raising this concern. We agree that unseen-instruction robustness alone does not establish cross-model protection. We therefore evaluate transfer between IP2P and SD3, and additionally to FLUX.2-klein-4B, a widely used open-source editor. The same protected images are reused without re-optimization, with the original instructions held fixed to isolate editor transfer. Each entry reports PSNR ↓ / LPIPS ↑ / ISR ↑.

Overall, SPIN achieves a higher average cross-model protection success rate, with a mean ISR of 72.00%, compared with 67.25% for SIFM[1], 65.00% for AdvDM[2], and 62.50% for SDS[3]. Despite differences in latent representations and denoising dynamics across editors, these results support the transferability of SPIN’s protection to unseen editing models without re-optimization.

Source → targetAdvDMSDSSIFM†SPIN
SD3 → IP2P17.82 / 0.3540 / 69%19.08 / 0.2986 / 65%16.72 / 0.4526 / 74%16.32 / 0.4434 / 78%
IP2P → SD319.36 / 0.3617 / 68%18.75 / 0.3752 / 69%18.19 / 0.4777 / 72%19.28 / 0.3802 / 68%
SD3 → FLUX21.21 / 0.1702 / 60%21.45 / 0.1760 / 55%20.80 / 0.1864 / 64%20.21 / 0.2019 / 70%
IP2P → FLUX20.14 / 0.2103 / 63%20.42 / 0.1965 / 61%21.33 / 0.1912 / 59%20.08 / 0.2145 / 72%
AVG19.63 / 0.2741 / 65.00%19.93 / 0.2616 / 62.50%19.26 / 0.3270 / 67.25%18.97 / 0.3100 / 72.00%

2.Comparison with more recent methods

Thank you for this suggestion. We now compare SPIN with both DiffusionGuard [4] and Universal Image Immunization [5] under the same editing protocol and an 8/255 perturbation budget. All experiments are conducted on single NVIDIA RTX 5090. The table reports ISR, output deviation from clean edits, and protection costs on unseen instructions. The baseline implementations are based on the methods described in their original papers.

MethodSD3 PSNR ↓SD3 LPIPS ↑SD3 ISR (%) ↑IP2P PSNR ↓IP2P LPIPS ↑IP2P ISR (%) ↑Offline trainingProtection time/image
DiffusionGuard [4]19.630.38956215.940.440780None17.33 s (SD3); 12.56 s (IP2P)
Universal Image Immunization [5]19.460.42046516.650.442978≈10.4 hNone
SPIN13.950.57048514.130.554391None16.01 s (SD3); 10.58 s (IP2P)

DiffusionGuard is also an optimization-based method and incurs a per-image cost comparable to SPIN, but achieves lower protection effectiveness. SPIN improves ISR over DiffusionGuard by 23% on SD3 and 11% on IP2P, with consistently better PSNR and LPIPS. Universal Image Immunization avoids per-image optimization by reusing a universal perturbation, but requires approximately 10.4 hours of offline training and still trails SPIN by 20% and 13% ISR, respectively. SPIN requires no offline training, and its per-image optimization is performed only once; the protected image can then be reused across editing instructions.

[1] Dong, S., Zhang, J., Zhao, G., Shan, S., & Chen, X. (2025). Semantic mismatch and perceptual degradation: A new perspective on image editing immunity. arXiv preprint arXiv:2512.14320.

[2] Liang, C., Wu, X., Hua, Y., Zhang, J., Xue, Y., Song, T., ... & Guan, H. (2023). Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. arXiv preprint arXiv:2302.04578.

[3] Xue, H., Liang, C., Wu, X., & Chen, Y. (2024, May). Toward effective protection against diffusion-based mimicry through score distillation. In International Conference on Learning Representations (Vol. 2024, pp. 25734-25758).

[4] Choi, W. J. S., Lee, K., Jeong, J., Xie, S., Shin, J., & Lee, K. (2025, May). Diffusionguard: A robust defense against malicious diffusion-based image editing. In International Conference on Learning Representations (Vol. 2025, pp. 27134-27180).

[5] Lee, C., Choi, D., Shin, S., Jeon, H. G., & Son, J. (2026, September). Universal image immunization against diffusion-based image editing via semantic injection. In European Conference on Computer Vision (pp. 463-481). Cham: Springer Nature Switzerland.

Thank you for pointing out the need to clarify the SD3 adaptations and optimization settings. MIST and SDS were independently re-optimized against each target editor, with model weights frozen.

For SD3, we adapt the VAE and DiT interfaces and replace the original noise-prediction residual with the native flow-velocity residual:

r=\hat v_\theta-(\epsilon-z),$$ where $z=E(x+\delta)$ is the scaled VAE latent. - **MIST:** We minimize $\|z-z_{\mathrm{target}}\|_2^2-10^4\operatorname{mean}(r^2)$, preserving its combined texture–semantic objective and backpropagating through both DiT and VAE. We retain the original target image and fusion weight $10^4$, without SD3-specific tuning. The target is resized to $512\times512$ and encoded by the target VAE; the texture loss uses sum reduction. Perturbations are initialized uniformly within the budget. - **SDS(-):** We inject the detached residual $r$ through the VAE and apply projected sign descent, preserving its update mechanism without DiT backpropagation or additional timestep weighting. Perturbations are initialized to zero. Both methods use **100 iterations**, an $\ell_\infty$ budget of $8/255$, a step size of $1/255$, and text/image guidance scales of **7.5/1.5**. To match SPIN’s instruction access, both methods optimize against the original editing instruction. Each protected image is reused for all five unseen instructions **without re-optimization**.

Thank you for answering my questions and sorry for late reply. The transfer experiments directly address my concern about reusing protected images on unseen editors, and the comparison with DiffusionGuard shows higher protection success rates at comparable per-image cost on both evaluated editors. The clarification of how MIST and SDS were adapted to SD3 also makes the original comparisons easier to assess. Since the authors addressed most of my concerns, I am raising my score from 2 to 3.

Thank you for your thoughtful feedback and for taking the time to reconsider our work. We are glad that our additional experiments and clarifications have addressed your concerns. We sincerely appreciate your constructive comments and updated assessment.

Reviewer_WS82OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an image immunity method called SPIN to protect images from unauthorized instruction-guided diffusion editing. Its core idea involves performing a single-step projection onto the predicted clean latent image and aggregating feedback from multiple random local editing trajectories. The results demonstrate that SPIN achieves robust protection with relatively low perturbation generation costs.

Critical strengths
  1. The paper proposes a relatively efficient method that uses one-step projection and stochastic local continuation to obtain information about potential editing outcomes.
  2. Strong empirical performance across multiple editing settings.
Critical weaknesses
  1. The proposed random neighborhoods are constructed around trajectories induced by optimization instructions, but it remains unclear why the local random variations around these trajectories are able to capture semantically distinct trajectories that are not induced by any observed instructions.
  2. Table 2 evaluates unseen instructions using perturbations optimized based on the original instructions. However, for instruction-agnostic baseline models such as AdvDM, MIST, and SDS, the perturbations themselves do not depend on the instructions. Could the authors clarify why the results for these approaches differ significantly between Table 1 and Table 2?
Questions for authors
  1. This paper does not provide a theoretical or reliability analysis of how the “one-step projection” objective predicts the failure of the final output under different images, trajectories, or instructions.
  2. How sensitive is SPIN to the specific stochastic trajectories sampled during optimization?
  3. How robust are these methods across different editing models?
  4. For instruction-agnostic baseline models such as AdvDM, MIST, and SDS, why are there such significant differences in Tables 1 and 2?
  5. The unseen-instruction evaluation uses five semantically disjoint instructions. How were these five instructions selected, and how representative are they of the broader instruction distribution?
  6. Is the ISR determined by the consensus of Kimi K2.5 and Qwen3-VL? What is the agreement rate between the models? How often do these two evaluators disagree?
5 comments

1.Rationale for Neighborhood Sampling and Generalization

Thank you for raising this concern. Because instructions are discrete, directly augmenting them or modeling their distribution requires additional choices about semantic validity and coverage. We instead augment continuous latent features, where Gaussian increments provide a tractable, differentiable neighborhood consistent with the editor's stochastic dynamics. Optimizing over these variations reduces dependence on a single instruction-induced trajectory, allowing transfer through the shared image-conditioning pathway without explicitly enumerating unseen instructions. Empirically, the comparison below between K=1K=1 and K=4K=4 at a fixed editor-specific neighborhood scale shows that neighborhood sampling improves protection under unseen instructions. These metrics demonstrate that perturbations optimized using the one-step objective remain effective beyond the optimization instruction.

EditorKKSeen PSNR ↓Seen LPIPS ↑Unseen PSNR ↓Unseen LPIPS ↑
SD3111.950.653214.570.5702
SD3411.200.673013.950.5704
InstructPix2Pix111.240.684815.510.5147
InstructPix2Pix410.930.697414.130.5543

2.Sensitivity to Stochastic Trajectory Sampling

How sensitive is SPIN to the specific stochastic trajectories sampled during optimization?

Thank you for asking about SPIN’s sensitivity to sampled trajectories. We tested three random sampling seeds on IP2P and SD3, with the same inputs and optimization settings. We report mean ± standard deviation of average scores across runs. Across both editors and instruction sets, standard deviations are below 0.085 dB for PSNR and 0.0035 for LPIPS. These results support low sensitivity to trajectory sampling on both original and unseen instructions in the tested settings.

Editing modelInstructionsPSNR ↓SSIM ↓VIFp ↓FSIM ↓LPIPS ↑
IP2POriginal10.985 ± 0.0450.3093 ± 0.00200.0653 ± 0.00110.6202 ± 0.00200.6928 ± 0.0032
IP2PUnseen13.943 ± 0.0330.4256 ± 0.00110.1068 ± 0.00100.6972 ± 0.00130.5571 ± 0.0019
SD3 UltraEditOriginal11.246 ± 0.0840.3459 ± 0.00460.0680 ± 0.00040.5831 ± 0.00430.6708 ± 0.0035
SD3 UltraEditUnseen14.000 ± 0.0520.4415 ± 0.00180.1037 ± 0.00130.6505 ± 0.00300.5691 ± 0.0015

3. Robustness across editing models

Thank you for raising this question. We evaluate cross-model robustness by transferring protected images between IP2P and SD3, and additionally to FLUX.2-klein-4B[1], an open-source image editor. The same protected images are reused without re-optimization.

Each entry reports PSNR ↓ / LPIPS ↑ / ISR ↑. The mean equally weights the four transfer directions.

Source → targetAdvDMSDSSIFM†SPIN
SD3 → IP2P17.82 / 0.3540 / 69%19.08 / 0.2986 / 65%16.72 / 0.4526 / 74%16.32 / 0.4434 / 78%
IP2P → SD319.36 / 0.3617 / 68%18.75 / 0.3752 / 69%18.19 / 0.4777 / 72%19.28 / 0.3802 / 68%
SD3 → FLUX21.21 / 0.1702 / 60%21.45 / 0.1760 / 55%20.80 / 0.1864 / 64%20.21 / 0.2019 / 70%
IP2P → FLUX20.14 / 0.2103 / 63%20.42 / 0.1965 / 61%21.33 / 0.1912 / 59%20.08 / 0.2145 / 72%
AVG19.63 / 0.2741 / 65%19.93 / 0.2616 / 62%19.26 / 0.3270 / 67%18.97 / 0.3100 / 72%

Cross-model performance depends on the source–target pair, as differences in latent representations and denoising dynamics can limit transfer. Nevertheless, SPIN achieves the highest ISR in three of the four directions, with a mean ISR of 72.00%, compared with 67.25% for SIFM, 65.00% for AdvDM, and 62.50% for SDS. These results support an overall advantage in cross-model protection under the tested settings without re-optimization.

[1]Black Forest Labs. FLUX.2 [klein]. https://huggingface.co/black-forest-labs/FLUX.2-klein-4B

4.Clarifying Instruction Conditioning in Tables 1 and 2

Thank you for highlighting this distinction. Our adaptations use the original editing instruction as the text condition during perturbation optimization. This makes the conditioning specific to the requested edit: MIST and SDS’s released implementations use generic descriptions such as "a painting" or "a photo," whereas we supply the original editing instruction. AdvDM’s original paper does not specify an exact conditioning string; our implementation likewise uses the original instruction in its denoising objective. This choice allows the protection objectives to account for the target editor’s predictions under the requested transformation. Each method is optimized through the corresponding frozen editor, retaining its core mechanism: prediction-error maximization for AdvDM, combined semantic and fixed-target texture objectives for MIST, and detached-residual updates through the VAE for SDS(-). Thus, the adaptation specifies which text condition guides optimization, rather than introducing text conditioning as a new mechanism.

Table 1 evaluates the instruction used during optimization, while Table 2 evaluates the same protected images under unseen instructions without re-optimization. The reported results therefore compare performance under the optimization instruction with cross-instruction transfer of these conditioned adaptations.

5. Selection and Coverage of Unseen Instructions

Selection. We used an LLM to generate five unseen instructions per image, prioritizing different editing intents from the seen instruction and plausible edits applicable to the source image. For instance, changing the same clothing to red instead of white remains a similar color-editing operation; we favor a different applicable operation. All five were excluded from perturbation optimization and evaluated on the same protected image.

Coverage. The full pool contains 2,500 image–instruction pairs across 500 images, spanning 11 editing categories. The distribution is shown as below.

Editing categoryInstructionsProportion
Attribute / surface changes66626.64%
Lighting / photographic appearance46018.40%
Weather / season / time of day42116.84%
Object addition36514.60%
Object removal1777.08%
Background / setting changes1757.00%
Artistic style / medium1154.60%
Object replacement632.52%
Text / symbol editing431.72%
Geometry / pose / quantity120.48%
Framing / canvas30.12%
Total2,500100.00%

Within-image instruction diversity. For each image, we compare each of its five unseen instructions with the instruction used to optimize the perturbation (seen–unseen). We also compare all ten possible pairs among the five unseen instructions (unseen–unseen). Across 500 images, these yield 2,500 and 5,000 pairs, respectively.

For each comparison, we report the percentage of pairs whose instructions belong to different editing categories in the table above, such as recoloring versus object removal.

ComparisonInstruction pairsPairs with different editing categories
Seen–unseen2,50090.36%
Unseen–unseen5,00086.18%

Thus, 90.36% of unseen instructions request a different editing category from the corresponding seen instruction, and 86.18% of unseen–unseen pairs differ in category. Additionally, 95.6% of images have unseen instructions covering at least three editing categories, showing that diversity is present within individual images.

6. Evaluator consensus and agreement

In our experements, ISR requires both Kimi K2.5 and Qwen3-VL to judge a sample as a defense success. Each evaluator assigns success when it detects semantic mismatch or severe quality degradation. Disagreements remain in the denominator and count as unsuccessful. Agreement on the final binary success labels is summarized below:

ModelSettingAgreement rate
SD3Seen93.50%
SD3Unseen88.20%
InstructPix2PixSeen95.50%
InstructPix2PixUnseen92.80%
AverageAll92.50%

These results indicate high agreement between the evaluators, while requiring both to agree on success ensures that disputed cases do not inflate the reported ISR.

Dear Reviewer,

We sincerely appreciate your time and effort in reviewing our work. Given the limited time available during the discussion period, we have briefly summarized the key points of our responses below for your convenience.

  1. Rationale for neighborhood sampling: Since text instructions are discrete, we model stochastic neighborhoods in the continuous feature space as a tractable alternative. We believe this is an empirically reasonable approach to improving generalization to unseen instructions, as supported by our experimental results.

  2. Sampling stability: Experiments across three random seeds demonstrate low sensitivity to sampled trajectories, with PSNR standard deviations below 0.085 dB and LPIPS standard deviations at or below 0.0035.

  3. Cross-model transfer: Without re-optimization, SPIN demonstrates overall superior cross-model transferability compared with the evaluated baselines.

  4. Baseline adaptation: We clarified that AdvDM, MIST, and SDS use the original editing instruction during optimization, ensuring consistent instruction conditioning.

  5. Unseen instruction diversity: Our evaluation includes 2,500 unseen image–instruction pairs across 11 editing categories, with 90.36% of seen–unseen pairs belonging to different categories.

  6. Evaluator reliability: The two VLM evaluators achieve 92.5% average agreement, and ISR conservatively requires both evaluators to agree on defense success.

We hope these additional experiments and clarifications have addressed your concerns. We would sincerely appreciate any further feedback you may have.

Thank you again for your valuable time and constructive comments.

Reviewer_7WFBOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPIN, an image immunization framework with one-step projection over local stochastic trajectory neighborhoods. Experiments on SD3 and InstructPix2Pix demonstrate strong performance across evaluation metrics under both seen and unseen editing instructions.

Critical strengths
  • The combination of single-step projection and stochastic neighborhood optimization provides an efficient approach to constructing protective perturbations without differentiating through the full denoising trajectory.

  • The proposed method demonstrates strong performance across the reported metrics on both diffusion editors.

Critical weaknesses
  • The Immunization Success Rate (ISR) is evaluated solely using consensus judgments from two vision-language models. However, VLM-based evaluations may not fully reflect human perceptions of editing failures and visual degradation. A human user study would provide stronger evidence for the effectiveness of the proposed protection method.

  • The experimental comparisons overlook recent optimization-free and universal image immunization approaches, such as [2] and [3]. Since SPIN requires iterative optimization for each image, comparisons with these methods in terms of protection effectiveness, computational efficiency, and scalability would provide a more comprehensive evaluation of its practical advantages.

  • The robustness of the proposed method is unexplored. Evaluating its effectiveness under common image transformations, such as JPEG compression, resizing, and cropping, as well as against perturbation purification methods such as IMPRESS [1], would provide stronger evidence of its practical robustness.

Questions for authors
  • Could the authors conduct a human user study to validate the VLM-based Immunization Success Rate (ISR) and assess whether the evaluation results align with human judgments?

  • Could the authors compare SPIN with recent optimization-free and universal image immunization methods [2, 3], particularly in terms of protection effectiveness, computational efficiency, and scalability?

  • Could the authors evaluate the robustness of SPIN under common image transformations, such as JPEG compression, resizing, and cropping, as well as against perturbation purification methods such as IMPRESS [1]?

[1] Bochuan Cao et al. IMPRESS: Evaluating the Resilience of Imperceptible Perturbations Against Unauthorized Data Usage in Diffusion-Based Generative AI. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 10657–10677, 2023.

[2] Tarik Can Ozden et al. DiffVax: Optimization-Free Image Immunization Against Diffusion-Based Editing. In Proceedings of the International Conference on Learning Representations (ICLR), pages 83–99, 2026.

[3] Chanhui Lee et al. Universal Image Immunization Against Diffusion-Based Image Editing via Semantic Injection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 463–481, 2026.

4 comments

1.Human evaluation

We agree that VLM-based evaluation should be complemented by human judgments. To assess whether VLM-based ISR reflects human perceptions of editing failures, we supplemented our evaluation with a human study involving three evaluators. We randomly sampled 200 per setting for each human evaluater. Each evaluater judged defense success based on semantic mismatch or severe visual degradation under the same setting of llms' judge. And we compare the human and VLM results use the same valid paired subset.

Editing modelSettingHuman ISRVLM ISRHuman–VLM agreement
SD3Seen98.0%90.0%92.0%
SD3Unseen96.0%88.0%85.0%
InstructPix2PixSeen96.0%92.0%92.0%
InstructPix2PixUnseen94.0%96.0%98.0%

The lower ISR for SD3 unseen suggests more conservative VLM judgments of failures such as visual degradation. Overall, the high human–VLM agreement supports using VLM consensus for automated evaluation.

2. Comparison with optimization-free and universal image immunization methods

We compare SPIN with DiffVax and Universal Image Imidentical test images, editing instructions, and an 8/2558/255 perturbation budget. We follow the baselines' settings for evaluation. All experiments are conducted on single NVIDIA RTX 5090. The table reports immunization success rate (ISR), output deviation from clean edits, and protection costs under unseen prompts.

MethodSD3IP2POffline timeProtection time/image
PSNR ↓LPIPS ↑ISR (%) ↑PSNR ↓LPIPS ↑ISR (%) ↑
DiffVax [1]22.070.20254623.220.161243≈22 h15.9 ms
Universal [2]19.460.42046516.650.442978≈10.4 hNone
DiffusionGuard [3]19.630.38956215.940.440780None17.33 s (SD3); 12.56 s (IP2P)
SPIN13.950.57048514.130.554391None16.01 s (SD3); 10.58 s (IP2P)

DiffVax [1] and Universal [2] offer fast protection without test-time optimization, but require substantial offline training (approximately 22 h and 10.4 h, respectively) and achieve lower ISR in our evaluation. SPIN achieves 85% ISR on SD3 and 91% on IP2P on unseen prompts, exceeding Universal by 20% and 13% points and DiffVax by 39% and 48% points, respectively, without training on any dataset. SPIN also outperforms optimization-based methods like DiffusionGuard, demonstrating stronger protection effectiveness under comparable computational budgets. Its time cost is incurred once for a specific image, and the protected image can be reused across editing instructions without re-optimization, making it practical for offline protection before publication.

3. Robustness of SPIN under common image transformations and purification

We assess robustness in a practical and challenging setting: protected images undergo common transformations or IMPRESS [4] purification before being edited with unseen instructions. The tables compare SPIN with SDS [5] and SIFM [6]. Lower PSNR and higher LPIPS indicate greater deviation from clean editing outputs, while higher ISR indicates stronger immunization. Best values are bolded.

Although SPIN is not explicitly optimized against these transformations or IMPRESS, it outperforms the baselines on most metrics, potentially due to its optimization over multiple stochastic denoising trajectories, which may reduce overfitting to a particular denoising path. For example, under JPEG (q = 70), IP2P ISR reaches 88%, compared with 79% for SDS and 80% for SIFM. Under IMPRESS, SD3 achieves 79% ISR, exceeding SDS (63%) and SIFM (76%). These results indicate that SPIN retains a degree of robustness to common image transformations and perturbation purification, even beyond the conditions used to generate the protection.

InstructPix2Pix — unseen prompts

Pre-processingSDS PSNR ↓SDS LPIPS ↑SDS ISR (%) ↑│SIFM PSNR ↓SIFM LPIPS ↑SIFM ISR (%) ↑│SPIN PSNR ↓SPIN LPIPS ↑SPIN ISR (%) ↑
None15.540.44684│15.620.48985│14.130.55491
JPEG (q = 70)17.980.34379│16.340.41880│15.410.45988
JPEG (q = 30)20.040.26958│18.160.31176│17.580.35780
Resize (256)19.280.30565│16.690.40682│17.050.39382
Crop (90% area)17.740.36076│17.720.38579│17.490.39081
IMPRESS [4]17.790.35876│16.980.45081│15.170.49386

SD3 UltraEdit — unseen prompts

Pre-processingSDS PSNR ↓SDS LPIPS ↑SDS ISR (%) ↑│SIFM PSNR ↓SIFM LPIPS ↑SIFM ISR (%) ↑│SPIN PSNR ↓SPIN LPIPS ↑SPIN ISR (%) ↑
None18.790.35569│17.810.48569│13.950.57085
JPEG (q = 30)21.430.26958│20.560.28860│20.430.29363
JPEG (q = 70)20.940.30162│18.960.36374│17.590.41678
Resize (256)22.170.24956│18.420.37475│18.280.38476
Crop (90% area)21.770.29354│19.520.27968│19.400.28968
IMPRESS [4]19.460.32163│18.060.36076│17.400.41879

[1] Tarik Can Ozden et al. DiffVax: Optimization-Free Image Immunization Against Diffusion-Based Editing. In Proceedings of the International Conference on Learning Representations (ICLR), pages 83–99, 2026.

[2] Chanhui Lee et al. Universal Image Immunization Against Diffusion-Based Image Editing via Semantic Injection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 463–481, 2026.

[3] June Suk Choi et al. DiffusionGuard: A Robust Defense Against Malicious Diffusion-based Image Editing. ICLR, 2025.

[4] Cao, B., Li, C., Wang, T., Jia, J., Li, B., & Chen, J. (2023). Impress: Evaluating the resilience of imperceptible perturbations against unauthorized data usage in diffusion-based generative ai. Advances in Neural Information Processing Systems, 36, 10657-10677.

[5] Xue, H., Liang, C., Wu, X., & Chen, Y. (2024, May). Toward effective protection against diffusion-based mimicry through score distillation. In International Conference on Learning Representations (Vol. 2024, pp. 25734-25758).

[6] Dong, S., Zhang, J., Zhao, G., Shan, S., & Chen, X. (2025). Semantic mismatch and perceptual degradation: A new perspective on image editing immunity. arXiv preprint arXiv:2512.14320.

I thank the authors for their response, the additional experiments address my concerns. I am increasing my score from 2 to 3, with the expectation that the authors will include these results in the final version of the paper.

Thank you for your constructive feedback and for taking the time to consider our response. We are glad that the additional experiments have addressed your concerns. We will include these results, together with the corresponding experimental details and discussion, in the final version of the paper.

Comments on the paper

1 comment

Comment: Dear Reviewers:

Thank you again for your thoughtful reviews. As the discussion phase approaches its end, we would like to check whether our response has addressed your main concerns. If any point remains unclear or if there are additional questions we could clarify, we would be happy to provide further clarification.

We greatly appreciate your time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.