acceptodds
Under review as a conference paper at ICLR 2027

Release, Don't Redraw: Instruction-Conditioned Release Control for Long-Horizon Multi-Turn Image Editing

Abstract

Real-world image editing is often multi-turn, yet recursively applying an editor can resynthesize unrelated content, causing identity drift, overwritten edits, and lost objects. We introduce ReleaseEdit, a training-free framework that inherits the accepted image and releases only content causally implicated by the current instruction. An instruction-conditioned contract separates proposal formation from acceptance. It uses evidence- and carrier-aware localized release for subject editing, prospective creation evidence with audited generation-as-commit for object addition, and boundary-aware entity transport for background replacement. Under localized acceptance, unreleased content is directly inherited rather than indiscriminately regenerated, and the accepted output becomes the sole persistent state for the next turn. On 400 six-turn sequences (2,400 edits), ReleaseEdit outperforms five baselines in average CLIP-T (0.3318), CLIP-Dir (0.1173), CLIP-I (0.9142), and DINO-I (0.8049), and remains best at Turn 6. It ranks first in 73.75% of 640 blinded comparisons, versus 18.91% for the strongest baseline. Ten-turn comparisons further illustrate reduced qualitative drift while preserving current-turn editability.

Primary area:
generative models
Keywords:
Multi-turn Image Editing, FLUX, Training-Free Editing, Diffusion Models, Release Field, Subject Consistency, Content Preservation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8EDAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents ReleaseEdit, a training-free framework for multi-turn image editing. Its main idea is to distinguish regions that may change from content that should be preserved, and to separate proposal generation from final acceptance. Different strategies are used for subject editing, object addition, and background replacement. The paper reports results on multi-turn editing sequences, but many qualitative examples are too blurry to support detailed visual inspection.

Critical strengths
  1. Multi-turn image editing is an important problem.
  2. The proposal–acceptance distinction is intuitive and potentially useful.
Critical weaknesses
  1. The technical novelty appears limited. The central idea is to modify target regions while preserving non-target regions. However, this method does not fully address the fundamental difficulty of multi-turn editing. If the same region is edited repeatedly, errors within that region can still accumulate and eventually collapse.
  2. The final image is obtained through mask-based compositing. This may introduce seams or inconsistent illumination near mask boundaries.

Minor:

  1. Figure 1 shows only five editing turns. Longer sequences would better demonstrate the performance.
Questions for authors
  1. How does ReleaseEdit compare with strong commercial multi-turn editors such as Nano Banana under the same instruction sequences?
  2. The method contains several task-dependent components and hyperparameters, but the paper does not provide sufficiently systematic robustness analysis. How sensitive is the method to the choices of hyperparameters?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 8EDA: Additional Evaluations and Method Clarifications

Thank you for your valuable suggestions and questions. We have carefully considered your feedback and provide our responses to each point below.

Q1. Comparison with Nano Banana

We have added a ten-turn comparison with Nano Banana, FLUX.2-klein-4B, and FLUX.1 Kontext [dev] in Figure 6, together with enlarged regional comparisons in Figure 7. All methods start from the same image, follow the same editing instruction sequence, and recursively edit their own outputs.

In this sequence, ReleaseEdit maintains stronger visual consistency than the two FLUX baselines, with less texture degradation and subject drift. At the fourth-turn background replacement, ReleaseEdit produces the requested pale-grey background while retaining the previously added yellow ball. Nano Banana leaves the background largely white and changes the ball into a tennis ball. The enlarged views in Figure 7 make these differences directly inspectable. We discuss the comparison in Section 4.4.

Q2. Hyperparameter sensitivity

We have added the default configurations for all three editing types in Table 3 (Appendix B, p. 14), qualitative sensitivity comparisons in Figure 5 (p. 9), and numerical diagnostics in Table 4 (Appendix E, p. 17).

The six-turn and ten-turn evaluations in Table 1 use fixed task-specific defaults shared across examples, without individual tuning. Common settings are FLUX.1-dev, 512×512 resolution, seed 42, and one candidate per instruction.

Figure 5 includes nearby settings for subject guidance g∈{3.0,3.5,4.0}g\in\{3.0,3.5,4.0\}, addition gain αadd∈{0.65,0.75,0.85,0.95}\alpha_{\mathrm{add}}\in\{0.65,0.75,0.85,0.95\}, and background width ω∈{0.015,0.020,0.025}\omega\in\{0.015,0.020,0.025\}. Each sweep fixes the input, prompts, observations, seed, sampling schedule, and acceptance rule.

The results expose the completion–preservation trade-off: increasing addition gain from 0.75 to 0.85 barely changes Edit scores while reducing local integrity, whereas nearby background widths produce similar preservation scores. Section 4.3 discusses these results. Diagnostic definitions and within-branch normalization are provided in Appendix E.

M1. Longer sequences and clearer qualitative examples

We have added quantitative evaluation on 400 ten-turn editing sequences in Table 1, a complete ten-turn comparison in Figure 6, and enlarged qualitative details in Figure 7.

Table 1. Final-turn results on 400 ten-turn editing sequences (p. 8). Each method recursively edits its own preceding output. All four metrics are higher-is-better.

MethodCLIP-T ↑CLIP-Dir ↑CLIP-I ↑DINO-I ↑
FlowEdit0.30870.04320.76480.5071
StableFlow0.29360.02840.77190.6153
MTC0.29450.05620.78370.5481
FireFlow0.29960.05270.81640.5318
RF-Inversion0.30690.05430.75260.4872
ReleaseEdit0.31780.11560.83410.6785

ReleaseEdit leads all five baselines across the four Turn-10 metrics. The additional figures expose the complete editing trajectory and local preservation details, addressing the concern about qualitative clarity.

W1. Technical novelty and error accumulation

ReleaseEdit separates the support used for proposal generation from the permission for acceptance. Subject proposals use broader semantic context with localized acceptance, while addition and background replacement use task-specific proposal controls and direct acceptance (Sections 3.4–3.6; Table 2).

For repeated editing, earlier requirements remain active unless superseded by the current instruction. This design reduces unintended changes while allowing new edits. The 400 ten-turn sequences in Table 1, together with Figures 6–7, provide evidence of sustained editability and reduced drift over the evaluated sequences.

W2. Mask compositing, seams, and illumination consistency

Acceptance is task-dependent. Addition and background replacement directly accept the generated proposal without a final pixel-space mask composite (Sections 3.5–3.6). Background masks control source-feature transport inside the transformer, allowing boundaries and surrounding appearance to form during generation.

Subject editing instead applies a continuous, scale-normalized release projection after context-aware proposal generation (Sections 3.3–3.4). Figure 7 provides unretouched enlarged views across all three edit types for inspecting boundaries, contact regions, and surrounding appearance.

Thank you for your response to my comments.

For technical novelty, the rebuttal largely restates the method as presented in the paper, so my concern is not addressed. I remain unconvinced that the proposed design is a sufficiently substantial technical contribution.

For comparison with Nano Banana, the examples in the rebuttal show cases where Nano Banana fails to follow the editing instruction and the proposed method produces a more accurate edit. However, the proposed method builds on the open-source FLUX and Qwen backbones. I therefore believe its overall editing accuracy is likely below that of Nano Banana, and a few selected failure cases do not change this. For a paper focused on multi-turn image editing, the more relevant question is whether the method better preserves image quality across successive edits. A convincing demonstration would be examples like Figure 6, where FLUX Kontext shows quality collapse after multiple editing turns. The authors would need to show that Nano Banana degrades in the similar way while the proposed method does not.

Given these remaining concerns, I will keep my original score.

Reviewer_Tz6qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ReleaseEdit, a training-free framework for recursive image editing with FLUX.1-dev. It separates the region used to form a subject-edit proposal from the region allowed to change in the accepted image, uses target-trajectory evidence and direct acceptance for object addition, and transports source representations to preserve foreground entities during background replacement. Automatic observations use Qwen2.5-VL-3B and SAM 2.

Critical strengths
  1. The task-dependent treatment of preservation is well motivated. An appearance edit, an added object with contact shadows, and a background replacement need different constraints, and the paper makes their acceptance rules explicit. In particular, permitting a broader subject proposal while restricting the final commit is a useful design distinction, and accepting an addition directly avoids the obvious failure of cutting a coherent new object with an overly narrow mask.

  2. The reported system-level gains are substantial and supported by more than source-image similarity. Table 1 reports Edit SR of 74.50 percent versus 52.25 percent for MTC, and Turn-6 CLIP direction alignment of 0.1315 versus 0.0565. ReleaseEdit also receives 472 of 640 first-place rankings in the blinded six-way study. Evaluating each method recursively on its own outputs, retaining failed attempts, and combining completion-oriented measures with human preference are meaningful strengths.

Critical weaknesses
  1. In Table 2, removing localized release reduces average DINO-I from 0.8049 to 0.5946, but removing the entire mechanism is not the same comparison as keeping the proposal solver, observations, and final projection fixed while changing the proposal field from the edit evidence to the broader carrier. Equation 5 makes this a particularly natural controlled comparison through the proposal-reach parameter. Likewise, the boundary/scale removal changes average DINO-I only from 0.8049 to 0.7983; because its intended benefit is specific to background edits, a background-only comparison against occupancy-only transport is needed to interpret it. The current results establish that the complete engineering system works better than the selected baselines, but leave uncertain which benefit requires the proposed interface rather than simpler mask and feature-transfer controls.
Questions for authors

Could the authors provide a matched single-field versus dual-field comparison with the same observations, sampler, and final commit rule, and a background-only occupancy-versus-boundary/scale comparison? These results would materially clarify the incremental contribution.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Tz6q: Matched Ablations and Extended Evaluation

Response to Reviewer Tz6q: Matched Ablations and Extended Evaluation

We thank the reviewer for the positive assessment of our work and the constructive suggestions on controlled ablations.

Q1. Matched single-field versus dual-field ablation

To complement the broader component removals in Table 2, we ran matched controls that isolate the proposal field while keeping the observations, sampler, prompts, seed, steps, guidance, and final acceptance rule fixed.

The single-field control uses the observed editing evidence E for both proposal generation and acceptance. The dual-field setting uses the broader carrier-derived field G for proposal generation while retaining the same E-based acceptance. The source-reference mixing coefficient is fixed at zero in both variants, removing the coupling between proposal extent and reference mixing.

Both variants use FLUX.1-dev at 512×512, 25 steps, guidance 3.5, and three paired seeds: 42, 43, and 44. The table reports means for two illustrative cases. Edit is the CLIP instruction gain; Integrity is the preservation diagnostic evaluated on the unedited subject region. Both are higher-is-better.

CaseSingle-field Edit ↑Dual-field Edit ↑Single-field Integrity ↑Dual-field Integrity ↑
Corgi recoloring0.00030.01540.99870.9452
Local mane recoloring−0.00030.01470.99990.9935

Dual-field proposal generation improves instruction response in all three paired seeds for both cases. For the local mane edit, the observed editing support covers approximately 10% of the horse carrier, so the proposal can use substantially broader subject context without expanding final acceptance. The dual-field setting maintains an unedited-region Integrity of 0.9935 in this case.

Q2. Matched occupancy-only versus boundary/scale-aware transport

We have added the matched comparison and an input-construction schematic to the SI, Appendix J, Figures J.1–J.4 (pp. 25–28), including enlarged foreground details, paired-seed diagnostics, and full-frame outputs.

To specifically test small-foreground preservation, the observed foreground occupies only 1.21% of the full 512×512 image in our synthetic scale stress test. Figure J.1 visualizes the fixed rescaling of the source image and its automatic foreground observation, together with the reduced foreground mask. Both policies share the input, observations, prompts, inversion, sampling settings, transport schedule, and three paired seeds. Both directly accept the generated proposal without a final pixel-space composite. Runtime audits verify that the full policy activates its scale-dependent residual branch.

Transport policyForeground Integrity ↑Foreground RGB MAE ↓
Occupancy-only0.239946.49
Boundary/scale-aware (ours)0.572220.94

Across all three paired seeds, the full policy improves foreground preservation, reducing mean foreground MAE by 55.0%. Figure J.2 directly shows the stronger retention of vehicle color and structure. Figure J.3 reports foreground preservation together with background instruction gain, while Figure J.4 provides every full-frame output. These controlled results support the added value of boundary/scale-aware transport for retaining small foregrounds during background editing, with the preservation benefit arising during generation rather than from output-level pixel compositing.

Additional evidence: ten-turn evaluation and commercial comparison

We have extended the evaluation to 400 ten-turn editing sequences, with each method recursively editing its own preceding output. Table 1 reports the final-turn results.

MethodCLIP-T ↑CLIP-Dir ↑CLIP-I ↑DINO-I ↑
FlowEdit0.30870.04320.76480.5071
StableFlow0.29360.02840.77190.6153
MTC0.29450.05620.78370.5481
FireFlow0.29960.05270.81640.5318
RF-Inversion0.30690.05430.75260.4872
ReleaseEdit0.31780.11560.83410.6785

ReleaseEdit leads all five baselines across the four Turn-10 metrics.

We also added a ten-turn comparison with Nano Banana, FLUX.2-klein-4B, and FLUX.1 Kontext [dev] in Figure 6, with enlarged details in Figure 7 and discussion in Section 4.4. All methods start from the same image, follow the same instruction sequence, and recursively edit their own outputs. In this sequence, ReleaseEdit maintains stronger visual consistency than the two FLUX baselines. At the fourth-turn background replacement, it produces the requested pale-grey background while retaining the previously added plain yellow ball; Nano Banana leaves the background largely white and changes the ball into a tennis ball. The enlarged views make these sequence-specific differences directly inspectable.

I appreciate the additional results and analysis. I have no further questions for the authors and recommend waek acceptance.

Reviewer_i54jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents ReleaseEdit, a training-free approach to reducing unintended changes during multi-turn image editing. It distinguishes subject appearance edits, object additions, and background replacements, and applies a different proposal and acceptance strategy to each. The central idea is to carry forward the previously accepted image while allowing the current instruction to change the content it requires. The paper evaluates the approach on mixed-operation editing sequences using comparisons with existing methods, component ablations, automatic metrics, and human preferences. It also provides illustrative longer sequences.

Critical strengths
  1. A well-motivated problem formulation. Multi-turn editing requires both compliance with the latest instruction and retention of earlier valid edits. Treating preservation as dependent on the requested operation, rather than imposing one uniform constraint, addresses a consequential limitation of recursive editing.
  2. Evidence from complementary forms of evaluation. The paper examines instruction following and preservation through quantitative comparisons, ablations, and human judgments. This breadth makes the reported benefit more persuasive than a claim based solely on selected visual examples.
Critical weaknesses
  1. The scope of the general and long-horizon claims exceeds the validation. The chosen operations cover important cases but leave deletion, geometric changes, and compound instructions largely untested. These limitations matter if the paper is read as a general solution to multi-turn editing, rather than a method for the particular operations evaluated.
  2. The paper appropriately distinguishes localized acceptance for subject edits from proposal-direct acceptance for addition and background replacement. A remaining empirical question is how reliably the latter two branches preserve previously accepted entities and edits without an output-level constraint. Branch-specific measurements of protected-content changes and audit-flagged failures would clarify the effectiveness and limits of this design choice.
  3. Eq. (4) is incorrectly written: the condition Hk(x)=1H_k(x)=1 should be stated explicitly as a condition for Ik(x)=Ik−1(x)I_k(x)=I_{k-1}(x), rather than appearing as a separate comma-separated expression. In Eq. (7), calign(t)c_{\mathrm{align}}(t) is described as an alignment correction, but no explicit definition of how it is computed is provided. Furthermore, the method section introduces many symbols, while the complete procedure is presented only as Algorithm 1 in the appendix, making the method unnecessarily difficult to follow. These presentation issues should be addressed.
Questions for authors

Please refer to the weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer i54j: Extended Evaluation and Method Clarifications

Thank you for your valuable suggestions and questions. We have carefully considered your feedback and address each point below.

W1. Scope of supported operations and longer-horizon validation

We have added quantitative evaluation on 400 ten-turn editing sequences in Table 1 (p. 8). Each method recursively edits its own preceding output. The final-turn results are reproduced below; all four metrics are higher-is-better.

MethodCLIP-T ↑CLIP-Dir ↑CLIP-I ↑DINO-I ↑
FlowEdit0.30870.04320.76480.5071
StableFlow0.29360.02840.77190.6153
MTC0.29450.05620.78370.5481
FireFlow0.29960.05270.81640.5318
RF-Inversion0.30690.05430.75260.4872
ReleaseEdit0.31780.11560.83410.6785

ReleaseEdit leads all five baselines across the four final-turn metrics. We have also added a complete ten-turn comparison and enlarged regional details in Figures 6–7 (p. 10).

Compound requests involving subject appearance editing, object addition, and background replacement can be explicitly decomposed into atomic instructions and executed sequentially through the existing multi-turn interface, with each accepted output becoming the next input.

Object removal can be incorporated as an additional task-specific proposal policy. We conducted an exploratory masked-inpainting test using the same frozen FLUX.1-dev backbone: the example largely removes the red bandana while retaining the other visible accessories. The comparison and configuration are included in Appendix H, Figure H.1 (p. 23). This demonstrates a compatible removal pathway, separate from the systematically evaluated ReleaseEdit branches and benchmark statistics.

W2. Preservation in the addition and background branches

We have added branch-specific results across the 400 ten-turn sequences in Table 5 (Appendix I, p. 24). The results are reproduced below:

MeasureObject additionBackground replacement
Evaluated turns1,5701,245
Current instruction completed1,512 (96.31%)1,227 (98.55%)
Prior content damaged or lost206 (13.12%)49 (3.94%)
Audit-flagged turns155 (9.87%)43 (3.45%)
Flagged turns confirmed problematic155/15543/43

Percentages use the corresponding branch total. Current-instruction completion and prior-content damage are separate outcomes and can overlap; completion alone is not a joint edit-and-preservation success rate. The final row reports confirmations among flagged cases, rather than detection recall.

Preservation is controlled during proposal formation: addition anchors the complement of prospective creation evidence to the source reference, while background replacement uses source-feature transport for foreground entities. Both directly accept the proposal (Sections 3.5–3.6). Audit records are diagnostic and do not modify the accepted output (Section 3.7). Figures 6–7 provide complementary qualitative evidence.

W3. Clarity of notation and presentation

We sincerely thank the reviewer for carefully reading our paper and pointing out these issues in typesetting, notation, and exposition. We have incorporated the following clarifications into Sections 3.2–3.4.

For Eq. (4), Section 3.3 (p. 5) now explicitly distinguishes protected pixels from release-weighted updates:

\begin{cases} I_{k-1}(x), & H_k(x)=1,\\\\ I_{k-1}(x)+A_k(x)[Q_k(x)-I_{k-1}(x)], & H_k(x)=0. \end{cases}$$ For **Eq. (7)**, **Section 3.4 (pp. 5–6)** now defines, for each packed latent token, $$e(t)=v_s(t)-v_t(t),\qquad r(t)=\dot z_{\mathrm{ref}}(t)-v_s(t),$$ $$c_{\mathrm{align}}(t)=\beta(t)e(t)$$ $$\beta(t)=\operatorname{clip}\left( \frac{\langle e(t),r(t)\rangle} {\max\left(\lVert e(t)\rVert_2^2,\varepsilon_{\mathrm{fp}}\right)}, 0,1 \right)$$ The inner product is over token channels, with $\beta(t)=0$ when $e(t)=0$. The source and target predictions are evaluated at $z_{\mathrm{ref}}(t)$ and $z_{\mathrm{ref}}(t)+h_k(t)$, respectively. The denominator floor is the smallest positive normal float32 value, not a tunable threshold. We have also rewritten **Section 3.2 (p. 4)** in execution order: identify the requested change and protected entities, obtain task-specific evidence and release roles, generate the proposal, apply the acceptance policy, and update the accepted state and audit record. **Algorithm 1, Appendix B (p. 14)** provides the equation-level procedure.
AuthorsOpenReview ↗

Polite Follow-up: Awaiting Your Response​​

Dear Reviewer i54j,

We hope this message finds you well. We greatly appreciate the time you have dedicated to evaluating our manuscript.

With the discussion phase coming to an end shortly, we wanted to send a polite reminder regarding our responses and the supplementary experiments. We truly hope our revisions have helped clarify your concerns.

If you still have any doubts or require additional information, please feel free to reach out at any time. We will do our best to provide a prompt response within the remaining time.

Best regards, The authors of Paper

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.