pith. sign in

arxiv: 2605.02417 · v2 · pith:TYGMOWYBnew · submitted 2026-05-04 · 💻 cs.CV

DirectEdit: Step-Level Accurate Inversion for Flow-Based Image Editing

Desong Yang , Mang Ye This is my paper

Pith reviewed 2026-07-01 00:37 UTC · model grok-4.3

classification 💻 cs.CV
keywords flow-based image editinginversionreconstruction fidelitytext-to-image modelsattention injectionnoise blendingtraining-free editing
0
0 comments X

The pith

DirectEdit removes reconstruction error in flow-based image editing by directly aligning forward paths without extra computations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to fix accumulated drift that arises when reconstruction and editing paths use mismatched noisy latents from different timesteps. It does so by aligning the forward paths inside the flow transformer itself, which produces exact reconstruction and allows safe feature sharing. This matters to a reader because existing training-free editing methods are limited by that drift, and removing it could make prompt-driven changes more reliable while keeping the same computational cost. The method adds attention feature injection plus multi-branch mask-guided noise blending to maintain editability alongside the new fidelity.

Core claim

DirectEdit eliminates the inherent reconstruction error without introducing additional neural function evaluations by directly aligning the forward paths in the flow transformer. This enables precise reconstruction and reliable feature sharing. A preservation mechanism based on attention feature injection and multi-branch mask-guided noise blending balances fidelity and editability, yielding superior performance across diverse scenarios compared with prior state-of-the-art approaches.

What carries the argument

Direct alignment of forward paths in the flow transformer, which removes timestep mismatch between reconstruction and editing paths and thereby carries the argument for error-free reconstruction and feature sharing.

Load-bearing premise

Directly aligning the forward paths will produce precise reconstruction and allow reliable feature sharing while preserving editability without side effects from the attention injection or noise blending steps.

What would settle it

Measure reconstruction error on a held-out set of images after applying DirectEdit; if error remains comparable to prior inversion methods rather than dropping near zero, the central alignment claim is falsified.

Figures

Figures reproduced from arXiv: 2605.02417 by Desong Yang, Mang Ye.

Figure 1
Figure 1. Figure 1: We present DirectEdit, a simple yet highly effective training-free method for flow-based image editing. Compared with existing inversion-based approaches, DirectEdit explicitly aligns the reconstruction and inversion trajectories and introduces an effective latent feature interaction mechanism, enabling step-level accurate reconstruction and precise background preservation. Extensive experiments across div… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of Inversion Methods. (a) Standard Euler inversion. Due to the accumulation of approximation errors between the reconstruction and inversion paths, it fails to accomplish successful reconstruction and editing. (b) Inversion via stepwise correction. Although error accumulation is mitigated, the persistence of step-level reconstruction errors results in the continuous injection of “drifted” featur… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of DirectEdit. Left: Direct Alignment for Accurate Inversion. By explicitly aligning with the inversion trajectory, we achieve step-level accurate reconstruction, thereby facilitating the extraction of ideal source image features. Right: Latent Feature Interaction. We further introduce a preservation mechanism that leverages noisy latents and attention features from the reconstruction path. This m… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with various editing methods. Our method demonstrates exceptional performance across a diverse range of editing tasks, outperforming prior state-of-the-art approaches in terms of both background preservation and text alignment. 4. Experiments 4.1. Setup Evaluation Datasets and Metrics. We evaluate our pro￾posed method and all baselines on the PIE-Bench dataset (Ju et al., 2023) for i… view at source ↗
Figure 5
Figure 5. Figure 5: Trade-off between CLIP similarity versus PSNR. DirectEdit achieves the optimal balance between editability (CLIP) and background preservation (PSNR) compared to other methods. Connected markers represent different hyperparameters. consistency with editing prompts, often resulting in under￾editing. In comparison, DirectEdit excels across diverse editing scenarios, achieving the best balance between pre￾serv… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of reconstruction errors across different inversion methods. DirectEdit demonstrates the lowest recon￾struction error among all compared approaches. still suffers from severe error accumulation. Stepwise Cor￾rection (Ju et al., 2023) employs a strategy that realigns the trajectory with the correct inversion path after each recon￾struction step; however, step-level errors remain pronounced. Build… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on attention feature injection steps in DirectEdit. Attention feature injection primarily influences the consistency between the edited region and the source image, where a greater number of injection steps results in higher similarity. 5. Conclusion In this paper, we present DirectEdit, a simple yet highly effective training-free framework for flow-based image edit￾ing. Unlike existing meth… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on attention feature injection steps in DirectEdit. Attention feature injection primarily influences the consistency between the edited region and the source image, where a greater number of injection steps results in higher similarity. significantly influences background fidelity, and the intro￾duction of multi-branch masks further enhances background preservation. Further ablation study re… view at source ↗
Figure 8
Figure 8. Figure 8: DirectEdit with Virtual Reconstruction. As shown in Algorithm 1, DirectEdit achieves precise reconstruction and facilitates drift-free feature interaction. Crucially, we observe that the editing process leverages features extracted from the reconstruction path. Given that the reconstruction path is explicitly aligned with the inversion trajectory in our framework, the computation of the reconstruction path… view at source ↗
Figure 9
Figure 9. Figure 9: System prompt for generating multi-branch mask. D. More Implementation Details In our experiments, we implement our method utilizing FLUX.1-dev (Labs, 2024) and SD3.5-medium (Esser et al., 2024) as the backbone models, respectively. For both architectures, the number of denoising steps is uniformly set to 30. We configure the Classifier-Free Guidance (CFG) (Ho & Salimans, 2022) scale to 1 for the inversion… view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative comparisons on PIE-Bench. E.2. More Ablation Study We conducted additional ablation studies on FLUX.1-dev to investigate the impact of various design choices on editing performance. Exp. ⃝1 employ the standard Euler method; due to the absence of editing components such as attention injection and noise blending, this configuration is equivalent to direct generation guided by the targ… view at source ↗
Figure 11
Figure 11. Figure 11: presents additional results of DirectEdit applied to high-resolution real-world images. As observed, DirectEdit achieves versatile image editing across diverse scenarios, capable of handling diverse local edits (e.g., object replacement, attribute modification, and fine-grained editing) as well as global style transfer (e.g., transforming photographs into painting or cartoon styles). Furthermore, by lever… view at source ↗
Figure 12
Figure 12. Figure 12: Failure case study. Our method shows inherent limitations when handling specific editing tasks (e.g., size change, spatial movement, viewpoint change, complex reasoning). 16 view at source ↗
read the original abstract

With recent advancements in large-scale pre-trained text-to-image (T2I) models, training-free image editing methods have demonstrated remarkable success. Typically, these methods involve adding noise to a clean image via an inversion process, followed by separate denoising steps for the reconstruction and editing paths during the forward process. However, since the reconstruction path is approximated using noisy latents from mismatched timesteps, existing methods inevitably suffer from accumulated drift, which fundamentally limits reconstruction fidelity. To address this challenge, we systematically analyze the inversion process within the flow transformer and propose DirectEdit, a simple yet effective editing method that eliminates the inherent reconstruction error without introducing additional neural function evaluations (NFEs). Unlike most prior works that attempt to rectify the inversion path, DirectEdit focuses on directly aligning the forward paths, enabling precise reconstruction and reliable feature sharing. Furthermore, we introduce a preservation mechanism based on attention feature injection and multi-branch mask-guided noise blending, which effectively balances fidelity and editability. Extensive experiments across diverse scenarios demonstrate that DirectEdit achieves efficient and accurate image editing, delivering superior performance that outperforms state-of-the-art methods. Code and examples are available at https://desongyang.github.io/Directedit.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes DirectEdit, a training-free method for text-to-image editing in flow-based models. It claims that directly aligning the forward paths in the flow transformer eliminates the inherent reconstruction error from mismatched timesteps in inversion-based approaches, without extra NFEs. A preservation mechanism using attention feature injection and multi-branch mask-guided noise blending is introduced to balance fidelity and editability, with experiments asserted to show superior performance over SOTA methods.

Significance. If the central claim holds, the work would address a key limitation in flow-model inversion for editing by providing a parameter-free alignment strategy that avoids accumulated drift. The public code and examples strengthen reproducibility and allow direct verification of the no-extra-NFE claim.

major comments (3)
  1. [Abstract] Abstract: the claim that 'extensive experiments across diverse scenarios demonstrate ... superior performance that outperforms state-of-the-art methods' supplies no quantitative metrics, baselines, or description of how reconstruction fidelity or editability were measured; this directly weakens the experimental support for the zero-error claim.
  2. [Method (preservation mechanism)] Preservation mechanism (attention feature injection + multi-branch mask-guided noise blending): no derivation or invariance argument is given showing that these operations leave the aligned forward-path velocity field unchanged at every timestep; because the central claim requires exact path preservation to eliminate drift, the absence of such analysis is load-bearing.
  3. [Method] The weakest assumption noted in the skeptic analysis—that the interventions preserve exact alignment without reintroducing timestep mismatch or feature drift—is not addressed by any consistency check or ablation that isolates the effect of the blending step on the flow trajectory.
minor comments (2)
  1. [Abstract] The abstract would be clearer if it briefly stated the key quantitative improvements (e.g., reconstruction PSNR or LPIPS deltas) rather than only qualitative assertions.
  2. [Method] Notation for the flow transformer velocity field and the exact definition of 'direct alignment' should be introduced with an equation early in the method section to avoid ambiguity when describing the preservation steps.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our central claims. We address each point below and will revise the manuscript accordingly to strengthen the experimental support and methodological justification.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'extensive experiments across diverse scenarios demonstrate ... superior performance that outperforms state-of-the-art methods' supplies no quantitative metrics, baselines, or description of how reconstruction fidelity or editability were measured; this directly weakens the experimental support for the zero-error claim.

    Authors: The abstract is a high-level summary; the full quantitative evaluation (PSNR/LPIPS for reconstruction fidelity, CLIP directional similarity for editability, and comparisons against inversion-based baselines such as Null-text Inversion and Prompt-to-Prompt) appears in Section 4 and Tables 1–3. We agree the abstract would better support the zero-error claim if it referenced these metrics and will revise it to include a concise statement of the evaluation protocol and key numerical improvements. revision: yes

  2. Referee: [Method (preservation mechanism)] Preservation mechanism (attention feature injection + multi-branch mask-guided noise blending): no derivation or invariance argument is given showing that these operations leave the aligned forward-path velocity field unchanged at every timestep; because the central claim requires exact path preservation to eliminate drift, the absence of such analysis is load-bearing.

    Authors: The operations are applied at identical timesteps to both paths and act as linear feature copies or convex combinations within the flow transformer; we will add a short invariance argument in Section 3.3 showing that, under the flow-matching ODE, these steps commute with the velocity prediction and therefore preserve the aligned trajectory. This addresses the load-bearing requirement for exact path preservation. revision: yes

  3. Referee: [Method] The weakest assumption noted in the skeptic analysis—that the interventions preserve exact alignment without reintroducing timestep mismatch or feature drift—is not addressed by any consistency check or ablation that isolates the effect of the blending step on the flow trajectory.

    Authors: We will insert a new ablation (Table X) that reports reconstruction error and per-timestep velocity alignment (measured by L2 difference between predicted velocities) with and without the mask-guided blending step, thereby isolating its contribution to trajectory fidelity. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method proposal is self-contained

full rationale

The paper proposes DirectEdit as a new algorithmic approach to align forward paths in flow transformers for inversion, supported by analysis and experiments rather than any mathematical derivation. No equations, fitted parameters, or self-citations are shown that reduce the central claims (zero reconstruction error via alignment, reliable feature sharing) to inputs by construction. The preservation mechanism (attention injection and mask-guided blending) is presented as an added component to trade off fidelity and editability, not as a tautological consequence of the alignment step. The result is therefore independent of the patterns that would indicate circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract introduces no explicit free parameters, axioms, or invented entities beyond standard components of flow-based T2I models.

pith-pipeline@v0.9.1-grok · 5732 in / 963 out tokens · 41140 ms · 2026-07-01T00:37:16.255826+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

4 extracted references

  1. [1]

    Examples: Replacing a dog with a cat, removing a person, changing a shirt's color to red, altering facial expression

    Local: Editing a specific object or its attributes. Examples: Replacing a dog with a cat, removing a person, changing a shirt's color to red, altering facial expression

  2. [2]

    Examples: Changing the setting from a street to a beach, changing the background color to white, moving the subject t o a forest

    Background: Changing the environment or background details while preserving the main subject. Examples: Changing the setting from a street to a beach, changing the background color to white, moving the subject t o a forest

  3. [3]

    Examples: Turning a photo into an oil painting, changing summer to winter, converting day to night, cyberpunk style t ransfer

    Global: Holistic changes affecting the entire image atmosphere or style. Examples: Turning a photo into an oil painting, changing summer to winter, converting day to night, cyberpunk style t ransfer

  4. [4]

    type": One of [

    Other: Structural changes, additions, or partial modifications in a specific non -object region. Examples: Adding a bird in the sky, adding glasses to a face, modifying a specific texture patch. Output a JSON object strictly with these fields: - "type": One of ["Local", "Background", "Global", "Other"]. - "bbox": [x1, y1, x2, y2] representing the bounding...

    2024