- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies protective perturbations against inversion-free text-guided image editing such as FlowEdit. It argues that existing protections ignore the editing trajectory, so their perturbations are either attenuated by later solver steps or act only on the final output, and proposes SourceShield, which maximizes the norm of the weighted sum of source-branch velocities over solver timesteps without knowing the target prompt and uses chunk-wise VJPs to reduce memory. Experiments optimize perturbations on SD3 and DreamShaper (DS) for 127 images and cover unseen prompts, cross-model transfer, and preprocessing.
Critical strengths
S1. The source-branch objective needs no target prompt, which the protector cannot know, and is memory-efficient to optimize. Because the Jacobian is additive over timesteps, the chunk-wise VJP cuts peak memory from 67.23 GB to 27.28 GB while per-image time only rises from 8.4 to 10.1 minutes.
S2. The white-box results on SD3 are strong. On unseen prompt pairs the average FID is 226.4 against 128.9 for the strongest baseline, and ImageReward (IR) is −0.47 against the best baseline's 0.90. Switching from maximization to minimization drops FID from 241.3 to 111.3, so the direction of the objective matters.
Critical weaknesses
W1. The persistent-disruption explanation is supported neither by the derivation nor by Fig. 2(b) Sec. 4.2 argues that since is subtracted in Eq. (6), amplifying it persistently perturbs the velocity difference. But the target branch sees the same perturbation. By Eq. (2) and , the target and source variables coincide at the first step, and if both velocities shift in the same direction their difference need not grow. In Fig. 2(b), SourceShield's stays around 0.9 to 1.7 for , roughly at or below the baselines, and rises sharply only from onward, reaching about 12.5 at the end. Since the metric is normalized by the clean , the late rise may partly reflect a shrinking denominator. The Step-Deviation variant in Table 5 changes two things at once (per-step vs. accumulated, deviation vs. norm), so it cannot isolate the effect of accumulation. Yet Secs. 1, 4.2 and 6 all rely on persistent disruption to explain why the method works.
W2. The premise that existing methods ignore the editing trajectory rests on a questionable categorization of prior work Like the source-branch objective, AdvDM [r1] and the SDS method of Xue et al. (2024) query the denoiser on noised versions of the protected image, need no target prompt, and compute gradients per timestep. They differ mainly in the objective built from these outputs and in how the gradient is computed. Yet the paper classifies Xue et al. (2024) as optimizing the final edited output (Sec. 1), does not cite AdvDM, and compares with neither. Output-stage methods are excluded entirely on the basis of Fig. 3, where Sec. 3.2 names PhotoGuard while the legend says DiffusionGuard, an early-stage method, and the paper does not say whether PhotoGuard's encoder or diffusion attack was used. Sec. 1 also states that such methods are "in principle not able to model the accumulated disruption effects", but a terminal loss backpropagated through the full editing process already accounts for every step's velocity difference, at a higher computational cost.
[r1] Adversarial Example Does Good: Preventing Painting Imitation from Diffusion Models via Adversarial Examples. ICML 2023.
Questions for authors
- According to App. D.3, the seed is fixed to 42 during optimization, and clean and protected images are edited with the same inference seed under default settings. The public FlowEdit configuration also defaults to seed 42. Does editing reuse the per-step noise realizations from optimization? If you have results with independent seeds, do they match Table 1?
- What are IR and CLIP-T for unprotected images and for random noise with the same budget? Under transfer to SD3 and FLUX, all methods have IR between 1.11 and 1.33, and DS→SD3 gives an FID of 86.7 against 81.3 to 84.9 for the baselines. On SD3, DiffPure lowers FID from 243.3 (Table 1, Prompt A) to 106.1 (Fig. 6), leaving PSNR and IR only 0.1 and 0.05 ahead of the strongest baseline. Without such a reference or confidence intervals, how do these results support the "superior transferability" and "resilience" claimed in the abstract? What parameters were used for the three preprocessing operations?
- The editor runs the editing model (Sec. 2.2), so the number of solver steps, the starting step and the guidance scales are theirs to choose. Fig. 7 gives FIDs of 97.5 and 156.0 when only the first or the last 10 steps are perturbed. How well does the protection hold when these settings change, or when the source prompt is shorter or more generic? Which editing method is used on DS, and if it is the consistency-model-based InfEdit, how is defined?
- The four baselines were designed for instruction-based editing, personalization and inpainting. How were they adapted to SD3 and DS? EditShield and VAC-LE perturb the VAE latent, which enters every step. Do they also count as early-stage disruptions that Fig. 2 describes as attenuated by later steps? If the input is encoded with the posterior mean, as in the official FlowEdit code, does VAC-LE, which targets the posterior variance, remain effective?
- Widely used open-source editors such as FLUX.1 Kontext and Qwen-Image-Edit are instruction-based and have no source branch. Does the perturbation transfer to these models?
- Minor points. The 41.5% and 115.1% in the abstract would be better reported as absolute differences, since IR crosses zero, and a single average hides that the FID gains on SD3 and DS (about 76% and 6%) differ greatly. The FID in Table 1 is computed on only 127 images per group, far fewer than the Inception feature dimension, so its estimate is unreliable and needs support from the per-image metrics before it serves as the headline result. All methods share the budget, and a PSNR of 28.60 vs. 28.40 in Table 3 does not support imperceptibility as a contribution. The full set in Fig. 7 has many more steps than the 10-step windows, so coverage and step count are confounded. I could not find the code. Figs. 9 and 10 are very blurry, Eqs. (1) and (9) omit the VAE encoder, the caption of Fig. 1 contains an incomplete sentence, and Table 5 and Fig. 5 contain typos.