acceptodds
Under review as a conference paper at ICLR 2027

DynaFill: Dynamic Manifold-Consistent Flow Sampling for Localized Image Inpainting without Inversion

Abstract

Image inpainting using pre-trained generative models typically relies on inversion—projecting the masked image back into the model’s noise space—which is computationally expensive and often struggles to preserve unmasked content while generating coherent structures in masked regions. In this paper, we introduce DynaFill, a novel inversion-free framework for localized image inpainting. Unlike existing inversion-based or fixed-anchor inversion-free editing methods, DynaFill is driven by a Dynamic Manifold-Consistent Sampling strategy. Specifically, we decouple the sampling trajectory from the static source prior by dynamically anchoring the noise perturbation to the current generation state rather than the corrupted original image. This allows the generative path to adaptively slide along the data manifold, ensuring that the filled content remains structurally faithful to the global context. To further enhance efficiency and locality, we integrate an enhanced ODE solver that accelerates sampling by nearly 2×, alongside a localized latent blending mechanism that strictly preserves unmasked pixels while enabling full generative freedom in masked holes. Extensive experimental results demonstrate that DynaFill achieves state-of-the-art performance in terms of structural coherence and texture fidelity, validating the effectiveness of the proposed method for inversion-free inpainting.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Inpainting, Rectified Flow, Inversion-Free Editing, Dynamic Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_X7xaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DynaFill, a local image restoration framework based on a pre-trained rectified flow model that does not require inversion, designed to balance the generation quality of missing regions, content preservation in unmasked regions, and sampling efficiency. Its core idea is to use the current generated state as a dynamic anchor in place of a fixed damaged source image, thereby adaptively constructing noise perturbations to reduce the dependence of the generation trajectory on the static source reference. Building on this, the method introduces a local adaptive fusion mechanism that combines static and dynamic anchors to preserve unmasked content and improve the continuity of the restoration boundaries; simultaneously, it employs a prediction-correction-based second-order ODE solver to enhance sampling efficiency. The authors report that the method achieves a 2.2× acceleration in sampling and demonstrates leading image restoration performance in both quantitative and qualitative evaluations.

Critical strengths
  1. The paper discusses content generation, preservation of unmasked regions and boundary continuity in inversion-free image restoration, which meets the practical requirements of local restoration.

  2. Dynamic anchors are based on sound design rationale, and the correspondence between the method components and the targets is relatively clear.

  3. The writing and logic are clear and easy to understand.

Critical weaknesses

I write the weaknesses and details in the comments on the AI Documentation For Reviewing.

Questions for authors
  1. How does the enhanced ODE solver used in the paper differ from other relevant work?

  2. The paper claims that the acceleration was only compared against its own results; has a comparison with other relevant work been considered?

Reviewer_HAV1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents DynaFill, a training-free and inversion-free method for localized image inpainting. It combines dynamic anchoring inside the mask, fixed source anchoring outside, and a predictor-corrector solver. Experiments report competitive inpainting quality and a 2.2x speedup over 40-step Euler sampling.

Critical strengths
  1. The problem setting is practically relevant: the masked region needs enough generative freedom, while everything outside must stay identical to the source image.

  2. Solver, dynamic anchor, and blending are each ablated separately, and the assembled method is probably useful in practice.

Critical weaknesses
  1. The headline claim of strict preservation outside the mask (Eq. (11)) is not established. Eq. (13) only blends anchors when building the noisy input, with no per-step projection or final compositing, and out-of-mask pixel error is never measured.

  2. The algorithm is underspecified: initial state, noise reuse, latent-space mask downsampling, prompts, CFG, and corrector inputs are all unclear, the text and figure of Eq. (6) disagree, and broken shared-noise cancellation gets no new analysis.

  3. The SOTA claim is not supported: six baselines promised vs. four reported, FlowEdit (the most relevant inversion-free method) never compared, and each baseline runs its own default backbone while DynaFill uses SD3 Medium. Full-image PSNR/SSIM/LPIPS are also dominated by the unmasked region, with no mask-only metrics, human eval, or error bars.

  4. The 2.2x speedup compares a 10-step predictor-corrector (two NFE per step) against 40-step Euler — not a matched-NFE comparison, and predictor-corrector is a standard scheme anyway.

  5. The manuscript contains many presentation and formatting errors. For example, words are repeatedly run together (e.g., "inpaintingthe," "imageis," and "trajectoryit"), citation links are surrounded by conspicuous colored boxes, Sec. 3.3 discusses a wolf-to-cow example but says that AMB generates a cake, and the reproducibility statement refers to an appendix that is not included.

Questions for authors
  1. Is the final output composited with the source image at the pixel level? If not, what are the maximum and mean pixel errors outside the mask and around its boundary?

  2. Could the authors provide complete pseudocode clarifying the initialization, noise reuse, mask downsampling, CFG computation, corrector inputs, and the description around Eq. (6)?

  3. Could the authors clarify all methods included in Table 1 and compare against a localized FlowEdit or fixed-anchor baseline under the same SD3 Medium checkpoint?

  4. How does the predictor-corrector solver compare with Euler or Heun under the same number of velocity evaluations? Could the authors also report mask-only, unmasked-region, and boundary metrics?

Reviewer_UY9BOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

DynaFill is a training-free, inversion-free framework for text-guided localized inpainting with a pretrained flow-matching model (SD3). It builds on FlowEdit, which integrates an ODE from the source latent using the difference between target- and source-conditioned velocities. It adds three components:

  • Dynamic Manifold Anchoring. The target branch's noisy input is built from the current ODE state instead of the fixed source latent.
  • Predictor–corrector (Heun) integration. This replaces Euler steps to reduce the number of sampling steps.
  • Localized Adaptive Manifold Blending. The noisy model input uses the static source anchor outside the mask and the dynamic anchor inside it.

The method is evaluated on 1,000 image–mask pairs from PIE-Bench, EditBench and COCO against Blended Diffusion, SD3 Inpainting, BrushNet and PowerPaint. It reports the best PSNR, SSIM, LPIPS and FID, with a CLIP-Score comparable to the baselines.

Critical strengths

S1. Inversion-free flow editing was designed for global editing and has no explicit mechanism to enforce locality. Addressing this for mask-constrained editing is a worthwhile problem.

S2. The method is simple, training-free and easy to adopt, provided it proves effective.

Critical weaknesses

W1. The core claim of dynamic anchoring is incorrect as stated. FlowEdit's target input is already a function of the current state ZtFEZ^{FE}_t. Substituting Z^tsrc=(1−t)Z0src+tNt\hat Z^{src}_t=(1-t)Z^{src}_0+tN_t into Eq. 5 gives

Z^ttar=ZtFE+t(Nt−Z0src),\hat Z^{tar}_t = Z^{FE}_t + t(N_t - Z^{src}_0),

whereas Eq. 6 gives

Z^ttar=(1−t)ZtDA+tNt=ZtDA+t(Nt−ZtDA).\hat Z^{tar}_t=(1-t)Z^{DA}_t+tN_t = Z^{DA}_t + t(N_t - Z^{DA}_t).

The change is therefore only a re-weighting of the current state, with the source-latent offset removed. It is not a move from a static anchor to a dynamic one.

The paper also criticizes anchoring ZtsrcZ^{src}_t to Z0srcZ^{src}_0, but then applies its change to ZttarZ^{tar}_t. The source branch stays anchored to Z0srcZ^{src}_0. As the paper notes, the two velocity fields now see different noisy inputs. This weakens the noise-cancellation argument inherited from FlowEdit, so Vtar−VsrcV^{tar}-V^{src} is no longer a pure edit direction. The "manifold-consistent / online manifold alignment" interpretation is an intuition with no analysis or measurement behind it.

W2. The blending section mischaracterizes prior work and overclaims.

  • Eq. 12 is wrong for Blended Latent Diffusion, which re-noises the blended background latent to the same noise level as the inpainted foreground. Eq. 12 uses the clean latent. With this corrected, Eq. 13 is essentially the Blended Latent Diffusion scheme, so the claimed contribution largely disappears.
  • "Strictly preserves unmasked pixels" is not guaranteed. Blending happens at the noise-construction stage, but the ODE still updates the whole latent.
  • No unmasked-region error or boundary-specific metric is reported to support the seam-free claim.

W3. The enhanced ODE integration lacks novelty and is not compared fairly. It is a standard Heun / predictor–corrector sampler. Table 2 compares 10 predictor–corrector steps, which need 2 evaluations each (about 20 evaluations), against 40-step Euler. The fair comparison is Euler at about 20 steps, plus other higher-order solvers. It is also odd that 4× more Euler steps gives 3.8 dB lower PSNR, which suggests the solver changes the effective edit strength, not only numerical accuracy. The "2.2× speedup" is a runtime comparison for one configuration, not an accuracy-matched one.

W4. Baselines and evaluation do not support the claim that inversion-free is better.

  • Missing methods. No other inversion-based or inversion-free method is compared. FlowEdit, the direct parent, is absent, as is the trivial locality fix of FlowEdit with a masked velocity.
  • Outdated baselines. Blended Diffusion, BrushNet and PowerPaint are SD1.5/SD2.1 U-Net methods run with default settings, while DynaFill uses an SD3 flow-matching DiT. This confounds backbone with method.
  • Inconsistent reporting. The text promises "six approaches" and discusses FlowChef and SD1.5 Inpaint, but Table 1 shows four baselines.
  • Preservation-biased metrics. PSNR, SSIM and LPIPS measure similarity to the source. The paper does not say whether metrics cover the full image or the masked region, and the benchmark masks are not described. DynaFill also uses a tuned target CFG of 12, while the baselines use defaults.
  • Task scope. The method needs a source prompt and an intact source image, and the examples are object swaps (wolf→cow, pizza→cake). The task is closer to mask-guided editing than to inpainting.

W5. The ablations are inconsistent with the paper's explanations. The paper says the fixed anchor "preserves excessive information from the original latent and limits the semantic transition". Yet the fixed anchor has the higher CLIP-Score (29.77 vs 29.48) and worse perceptual scores. The same pattern appears in the solver ablation (CLIP 30.35 vs 29.48), which suggests a preservation-versus-edit-strength trade-off, not a strict improvement. Some differences are tiny (e.g., FID 68.13 vs 68.04) and no variance is reported. The text cites a sofa→chair example that does not appear in Fig. 4, which shows umbrella→parasol instead.

W6. The paper reads as incomplete.

  • The explanation of inversion-based and inversion-free editing is limited and confusing.
  • There is no algorithm section, and implementation details are sparse.
  • The FlowEdit notation (ZtFEZ^{FE}_t, ZtsrcZ^{src}_t, ZttarZ^{tar}_t) is copied without clear explanation, and FlowEdit is not credited where it is introduced, which makes these terms confusing.
  • There are minor typos (e.g., lines 31, 41, 64).
Questions for authors
  1. Given the algebra in W1, what does Eq. 6 do beyond FlowEdit's target input, and why does it help when the source branch is still anchored to Z0srcZ^{src}_0?
  2. Can you compare against FlowEdit, FlowEdit with a masked velocity, and at least one inversion-based method, all on the same SD3 backbone?
  3. How does Eq. 13 differ from Blended Latent Diffusion once the background is re-noised correctly?
  4. Can you compare Euler at about 20 steps against 10-step Heun, and against other higher-order solvers, at equal evaluation count?
  5. Can you report background-only error, boundary-region metrics and masked-region CLIP, and plot preservation versus CLIP as edit strength varies?
  6. How were the source prompts, masks and CFG values chosen for DynaFill and for the baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.