acceptodds
Under review as a conference paper at ICLR 2027

When Text-to-Image Helps Editing: The Effects of Conditioning During Denoising

Abstract

Unified models are trained for both instruction-based image editing and text-to-image (T2I) generation, but standard editing pipelines keep source-image conditioning throughout denoising. We ask whether editing can benefit from T2I, and study how the effects of conditioning vary across edits and denoising stages. In pure editing, source attention declines for some edits over the sampling trajectory. This observation led us to task switching, which lets the model draw on its T2I capabilities. Across three unified editors and four benchmarks, switching to the T2I task for bounded intervals improves edit quality, while mean perceptual preservation remains close to pure editing across all three models. Unified editors therefore benefit from using both conditioning modes they are trained for, and the timing of the switch sets the balance between quality and preservation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
instruction-based image editing, text-to-image generation, unified generative models, diffusion models, inference-time conditioning

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_6BqAOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem that persistent source-image conditioning may limit editing quality in unified image generation and editing models. The authors observe that different editing tasks exhibit varying degrees of dependence on the source image and that source-image attention changes dynamically throughout the denoising process. Based on this observation, the paper proposes temporarily switching to the T2I mode during certain denoising steps and then restoring editing conditioning to improve editing quality while preserving the original image content as much as possible. Experiments show that the proposed method improves editing quality across three unified models and four benchmarks while maintaining similar average perceptual preservation scores, although it introduces some pixel-level preservation loss.

Critical strengths
  1. The paper presents an interesting observation that different image editing tasks exhibit varying degrees of dependence on the source image, and that source-image attention follows different temporal patterns throughout the denoising process. This observation provides valuable insights into the conditioning mechanisms of unified image generation and editing models and offers an intuitive motivation for the proposed task-switching strategy.

  2. The paper provides systematic and extensive experimental analyses, covering different conditioning configurations, task-switching intervals and durations, guidance strategies, and the trade-off between editing quality and source-image preservation. Furthermore, the authors evaluate the proposed method across multiple models and benchmarks, providing convincing empirical support for its effectiveness.

  3. The paper presents a clear and well-motivated research problem, highlighting how persistent source-image conditioning may restrict the editing capabilities of unified models.

Critical weaknesses
  1. The paper's motivation is based on the dynamic changes in source-image attention during a 50-step denoising process and the potential redundancy of source-image conditioning at certain stages. However, diffusion distillation methods such as DMD[1] and DMD2[2] have demonstrated that conventional multi-step denoising can be compressed into one-step or few-step generation, and related works have also applied few-step sampling to image editing tasks. The phenomenon observed in this paper may not exist in one-step or few-step editing processes, which limited the paper's motivation.

  2. Temporarily removing source-image conditioning and introducing T2I generation during denoising may lead to unintended changes unrelated to the editing instruction. For example, in the second example of Figure 5, there appear to be unnecessary grass details around the fox, while in the third example, suspicious shadow changes can be observed near the bookshelf. These observations suggest that even after restoring editing conditioning, unintended changes introduced during the T2I phase may not be fully corrected.

  3. The proposed method relies on an additional VLM to generate a target caption based on the source image and editing instruction. However, this process may introduce incorrect semantic information or overemphasize details unrelated to the editing objective. Since source-image conditioning is completely removed during the T2I phase, such inaccurate descriptions may further lead to unintended content generation, and restoring editing conditioning in the subsequent stage may not necessarily correct these errors.

  4. Although the paper observes that source-image attention varies dynamically throughout the denoising process, the proposed method still relies on a fixed switching interval determined in advance through empirical sweeps. Specifically, they uniformly adopt [10,16) as the switching interval, while the attention-based selective switching strategy only determines whether to switch, rather than adaptively deciding when to start and end the T2I phase for different images and editing tasks. However, different images and editing tasks may exhibit substantially different dependencies on source-image conditioning, and a fixed interval may not be suitable for all cases. Furthermore, Table 10 shows that the optimal switching intervals vary across models and evaluation metrics. For example, BAGEL achieves its best GEdit overall score with [35,45) rather than [10,16). Therefore, this empirically determined fixed switching strategy may limit the adaptability and generalizability of the method. Further investigation into dynamically determining switching times based on source-image attention would be valuable.

  5. Although the paper discusses related works such as HiDream-E1[3], SaaS[4], GRAG[5], and RvR[6], the experiments mainly focus on internal ablation studies of different conditioning configurations, switching intervals, and guidance strategies, without direct comparisons with existing methods. In particular, HiDream-E1 has already adopted a staged editing-to-generation refinement approach, while SaaS and GRAG also improve editing quality by adjusting the relative influence of the source image and editing instruction. Therefore, the current experiments are insufficient to fully demonstrate the advantages of the proposed task-switching strategy over these existing approaches. I suggest that the authors include comparisons with relevant baselines under comparable model settings and computational budgets, or further explain why such comparisons are technically infeasible.

[1] One-step Diffusion with Distribution Matching Distillation

[2] Improved Distribution Matching Distillation for Fast Image Synthesis

[3] HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer

[4] SaaS: Self-Adaptive Attention Scaling

[5] Group-Relative Attention Guidance for Image Editing

[6] Enlarging Modification Space Boosts Image Refinement in Unified Multimodal Models

Questions for authors

Please see my weakness.

5 comments
AuthorsOpenReview ↗

Official Response to Reviewer 6BqA

We appreciate the reviewer’s recognition of our study’s motivation, its insights into source-image conditioning, and the breadth of the experimental analysis. We address each of your concerns below.

W1. Relevance to few-step editing

Distillation and our study address complementary questions: reducing sampling cost and improving editing through conditioning. We show that changing conditioning along the denoising trajectory improves editing quality. The resulting editor could then serve as a teacher for distillation, without requiring the student to switch as well. Few-step generation therefore does not remove the motivation for studying multistep trajectories. Whether distillation preserves these gains is an interesting direction for future work.

W2. Unintended changes

Yes, some minor details can change, including the shadows you noted and the candles on the table. Fig. 5 shows that returning to editing reduces these changes compared to a one-way switch, although it does not restore every source detail. The T2I interval gives the model more freedom to execute the requested edit. When pure editing fails to make the intended change, improved fidelity can be worth a small loss of unrelated detail. We report the preservation costs in §4.4 and examine how interval placement and duration affect this balance.

Our new paired TOST tests comparing fixed [10,16) with pure editing support equivalent mean GPT preservation within ±0.25 points on a 0–10 scale for all 3 models (results). Our human study also shows that GPT largely agrees with human preservation judgments. Together with the fidelity and naturalness gains reported in the paper, this evidence supports that T2I intervals can improve editing quality without a substantial loss of perceptual preservation.

W3. Target-caption reliability

Using a language model to improve generation prompts is established practice, including in DALL·E 3. Such rewriting can introduce errors, but its benefits outweigh these risks, as our E* results demonstrate. In our switching setup, the generated caption guides only a small part of the trajectory, which limits the influence of potential caption errors. We discuss caption errors in §5 as a potential failure mode, but have not noticed them as a recurring issue. Unintended content generation is covered by our preservation evaluations.

W4. Adaptive switching times

We had also explored whether attention statistics available during denoising could guide adaptive interval selection without additional forward passes. The signals we tested were insufficient for reliably choosing interval boundaries, even in a simpler classification task choosing among predefined intervals placed throughout the trajectory. Online exit selection also poses a specific challenge: source attention is unavailable during the T2I interval because the source image is absent. We therefore report timing sweeps and use attention for the simpler decision of whether to apply a fixed interval. We agree that adaptive timing is a valuable goal, but achieving it without increasing denoising latency proved difficult. Using editing–T2I prediction disagreement could help, but evaluating both predictions requires additional forward passes, making this a promising inference-time scaling direction beyond the scope of our current work. Regarding BAGEL’s later optimum, our extended Fig. 3 suggests an explanation for why preferred interval placement differs across models.

W5. Comparisons with related methods

HiDream-E1 switches from editor to generator weights. Our E→T2I mode already tests the same editing-to-generation handoff while keeping a unified model’s weights fixed. RvR targets T2I refinement rather than instruction-based editing and requires additional training. SaaS selects layers through OmniGen-specific perturbation analysis; GRAG uses text tokens for Kontext but source-image tokens for Qwen-Edit/Step1X-Edit. Neither specifies these choices for our backbones. We therefore cannot compare these methods directly: RvR requires additional training, while SaaS and GRAG require ambiguous adaptations to our models.

Thank you for the detailed response. While I appreciate the authors' clarifications, several of my main concerns remain insufficiently addressed, particularly due to the lack of additional experimental evidence. I therefore maintain my original score of 2.

Thank you for following up! We’d like to highlight the new evidence provided in our response:

  • Preservation equivalence tests [link]
  • A human study assessing human–judge agreement [link]
  • Fig. 3 extended to all 3 models, addressing differences in conditioning behavior across models [link]

Could you clarify what additional experimental evidence would help address your remaining doubts about our contributions?

Reviewer_4675OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies switching a unified model between source-conditioned editing and caption-only text-to-image generation within a single denoising trajectory. Source-attention and prediction-disagreement diagnostics motivate a bounded text-to-image interval, followed by a return to editing. Using unchanged weights, the authors compare image/text conditioning controls, interval placements, continuous guidance, and an attention-based selective rule across FLUX2-Klein 4B and 9B, BAGEL, and four editing benchmarks. A fixed interval from step 10 to step 16 improves the reported quality scores while reducing pixel-level preservation and leaving the model-judged preservation means relatively close to baseline.

Critical strengths
  1. This paper separates removing the source image from replacing the instruction, includes improved instructions generated with the same VLM-call budget, and compares a return to editing with a one-way handoff. The authors also acknowledge the HiDream-E1 precedent. These controls provide evidence about when source conditioning helps and why caption-only generation can improve perceptual quality.

  2. The effect transfers beyond one editor and one benchmark. In Table 2, fixed switching raises GIE-Bench functional correctness from 73.40 to 76.46 percent on FLUX2-Klein 4B, from 77.94 to 81.28 on 9B, and from 67.16 to 72.38 on BAGEL. The paper reports preservation costs instead of hiding them, includes paired source-clustered comparisons, and tests alternative captioners and an improved-instruction baseline. This breadth makes the finding informative even though switching itself is simple.

Critical weaknesses
  1. The claim that perceptual preservation is essentially unchanged is stronger than the analysis establishes. Table 16 reports a fixed-interval GPT-preservation change of -0.158 on 9B with a confidence interval from -0.338 to +0.023, and -0.075 on 4B with an interval from -0.246 to +0.096. These nonsignificant differences do not establish equivalence within a practically acceptable margin. Meanwhile, the corresponding SSIM losses are statistically clear, including -0.0391 on 9B. The paper correctly acknowledges that its model judge has not been validated against human judgments here, so attributing the discrepancy primarily to imperceptible shifts remains an interpretation. The evidence supports improved quality with a measured preservation trade-off and small mean judge-score changes. It might be better to include a human validation step.

  2. The selection procedure limits the strength of the generalization claims, particularly for attention-based routing. Section 4.1 chooses the featured interval after examining 4B sweeps across all four benchmarks. The 9B distribution-based transfer and the other benchmarks provide useful additional evidence, but the 4B routing advantage is not an independent test of selection quality. A held-out source split or cross-validation is needed to determine how much of that advantage survives selection, and the exploratory interval sweep should be distinguished from a frozen-policy evaluation.

Questions for authors
  1. What preservation loss would count as practically negligible on the GPT scale, and do the paired intervals support equivalence within that margin? A human comparison focused on unintended changes would materially strengthen the interpretation of the SSIM/judge disagreement.

  2. Can the interval and selective thresholds be frozen on a disjoint development subset and evaluated on unseen source images, including the comparison against preservation-matched random routing?

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer 4675

We thank the reviewer for recognizing the strength of our conditioning controls, the transfer across models and benchmarks, and our reporting of preservation costs. We are pleased to address the points you raised.

W1/Q1. Preservation equivalence and human evaluation

Following your suggestion, we added paired TOST tests and a human preservation study. Using the ±0.25-point margin suggested during review, all 3 models support equivalence in mean judged preservation on the 0–10 GPT scale after Holm correction. The full table is in our preservation follow-up.

To assess practical impact directly, we conducted a blind side-by-side study on 100 GIE editing samples for FLUX 4B, with 3 human ratings per sample. GPT largely agrees with human preferences, especially for large differences. When slight differences count as ties, humans rated preservation with [10,16) as equal to or better than pure editing on 74% of samples. The human study results include all ratings and the breakdown by how noticeable the changes were.

W2/Q2. Interval selection and attention-based routing

Our interval sweeps characterize how switching effects vary with placement and duration, with gains across multiple intervals. Table 2 features [10,16) to keep one configuration across all metrics, avoiding results from different intervals in the same row. A held-out source split or cross-validation would evaluate a rule for choosing intervals on new data. Developing such a rule is outside the scope of this study.

The GEdit comparison was a sanity check of our attention-based routing rule against random routing. To test whether its advantage survives beyond the data used to select the thresholds, we now compare the same thresholds, kept fixed, against random routing on GIE-Bench. Following Appendix E, we match random routing’s expected mean SSIM outside the edit mask to selective routing’s SSIM (FLUX 4B, 1,079 samples).

RoutingPreservation (SSIM)Functional correctness (%)
Pure editing0.80573.401
Attention rule (79% switched)0.78677.108
Random, preservation-matched (p = 0.78)0.78675.849

The routing advantage persists on GIE-Bench: 1.26 percentage points at equal mean SSIM, without retuning the selective thresholds.

Thanks for the response and new experiments~

The paired TOST tests, the human evaluation, and the out-of-sample routing check on GIE-Bench against SSIM-matched random routing largely address my questions on preservation and threshold transfer.

Since interval selection itself remains heuristic and the out-of-sample routing gain is modest, I will keep my rating of 3 and encourage the authors to incorporate these results into this paper.

Thank you for your feedback and for considering the additional experiments! We’re glad they addressed your questions, and we will add these results to the final version of the paper.

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

Unified models can do both instruction-based editing (source image + instruction) and text-to-image generation (caption only), but editing pipelines keep the source image conditioned at every denoising step. The paper asks whether an edit can profit from switching to the model's other mode for part of the trajectory. The diagnosis comes first. On FLUX2-Klein 4B, over 50-step pure-editing trajectories, source-attention mass depends on the edit and generally decays with denoising, while local edits keep high attention late (Figures 2–3a). Comparing the two modes' velocity predictions at the same latent state, cos(v_E, v_T2I) is low early, comparatively stable in the middle, and diverges again late for local edits (Figure 3b–c). That suggests a middle window where a switch disturbs least. The intervention: for a bounded interval [k1, k2), drop the source image, replace the editing instruction with a VLM-generated target caption, keep denoising the same latent, then restore editing. Weights are untouched, and the FLUX models still do two denoiser passes per step (Table 11); the extra cost is one caption call. Controls separate the image from the text: E* (improved instruction, source kept), caption with the source kept, editing instruction with the image removed, one-way E→T2I versus E→T2I→E. Results with a single fixed interval [10, 16) applied to every edit: all six quality scores go up on all three models. FLUX2-Klein 4B: GEdit SC 7.72→8.08, PQ 7.46→7.90, O 7.11→7.67; ImgEdit 3.74→3.99; MagicBrush CLIP-T 0.317→0.322; GIE-Bench 73.40→76.46. FLUX2-Klein 9B: SC 8.38→8.55, O 7.74→7.96, ImgEdit 4.02→4.23, GIE 77.94→81.28. BAGEL: SC 7.68→7.87, O 6.64→6.81, GIE 67.16→72.38. A GPT-4.1 preservation judge (GEdit's criterion with the fidelity part removed) moves by at most 0.16 out of 10 and is not significant. SSIM outside the mask and Unmasked-CLIP do drop significantly (4B 0.805→0.781, 9B 0.939→0.899, BAGEL 0.960→0.952). The paper also shows that removing the image is what unlocks perceptual quality while the caption keeps the scene on target, that an improved instruction is a strong competitor, and that a selective rule based on source attention (switching 68–69% of GEdit edits) recovers part of the SSIM loss. A multimodal-guidance blend does better on consistency and preservation, but at four denoiser passes instead of two.

Critical strengths

The chain from diagnosis to intervention to control is well done. The timing isn't a hyperparameter sweep: it comes from measured attention decay and from the agreement between v_E and v_T2I, and then the intervention is tested against controls that could have killed the story (image removed with the original instruction, caption with image kept, E*, one-way versus interval, continuous guidance). That's how an observation should be turned into a method. The image-versus-text decomposition is a real result. Table 1 and Figures 6–7 show the gain is specific to removing the source image, that the caption is what prevents semantic drift, and that a better instruction raises semantic consistency without the perceptual gain. This closes a confound that a "better prompt" baseline would leave open. The empirical base is broad and the comparison set is honest. Three unified editors including a non-FLUX architecture (BAGEL), four benchmarks, six quality metrics, and controls that include the strongest prompt-only competitor and a continuous-guidance alternative at higher cost. The paper reports the SSIM losses instead of hiding them, and it compares the selective rule against preservation-matched random routing (Table 15), which most papers skip.

Critical weaknesses

1. The ranking against E* flips with model and judge, and the quality significance tests never appear. Against pure editing, the improvements are real (Table 2). Against E*, it depends:

  • 4B, GPT-4.1 (Tables 5 and 10): interval O = 7.67 vs E* 7.54, but E* wins SC (8.17 vs 8.08) and has better SSIM (0.799 vs 0.781) and PSNR/MSE.
  • 9B, GPT-4.1 (Table 10): E* wins overall — SC 8.79, O 8.10 vs the best T2I interval's 8.55, 7.96. T2I wins PQ (8.22 vs 7.80).
  • 4B, Qwen3.5 judge (Table 8): E* reaches best O = 7.80 vs 7.69 for the best T2I interval. That's a legitimate finding — switching is one point on a Pareto curve, not a free win — but the abstract states it as a general improvement. And Section 4.1 promises "two-sided Wilcoxon signed-rank tests and exact McNemar tests for binary correctness on matched per-edit records", yet no quality test results appear anywhere in the paper. Table 16 is preservation only. The main quality claim currently rests on point estimates and baseline three-seed std. 2. Preservation is argued from non-significance and a modified judge, while the pixel metrics drop significantly. The selling point is "quality up, preservation essentially unchanged". The evidence is a GPT-4.1 judge using GEdit's preservation criterion with the fidelity part removed — a modification not validated against humans — plus non-significant paired differences:
  • 4B fixed [10,16): GPT-preservation Δ = −0.075, CI [−0.246, +0.096], Holm p = 0.946.
  • 9B fixed: Δ = −0.158, CI [−0.338, +0.023], p = 0.127.
  • BAGEL fixed: Δ = −0.044, p = 1.000. A wide CI containing zero isn't evidence of equivalence. The 9B estimate isn't obviously within any reasonable margin, and there's no TOST with a pre-specified margin. Meanwhile SSIM outside the mask drops significantly on every model (4B −0.0246, 9B −0.0391, BAGEL −0.0082, all p < 0.001), and the raw GIE-Bench numbers aren't marginal: at [10,16) PSNR falls 20.88→18.45 dB and MSE rises 1882→2570 (Table 3); at [10,20) SSIM is 0.684 (Table 14); a one-way switch is catastrophic (SSIM 0.54, PSNR 9.21, MSE 8892 at k = 5). The argument that SSIM is hypersensitive to small translations may well be right, but "our modified judge says it's fine", with no human study and no judge–human agreement, is thin support for the central claim. Note also that the same model family (GPT-4.1) judges both quality and preservation, so their errors may be correlated. 3. The interval and the selective thresholds are chosen on the benchmark they're reported on, and BAGEL doesn't fit the story. [10,16) comes from 4B GEdit sweeps and is then applied unchanged to 9B, BAGEL and three other benchmarks; 9B/BAGEL were tested on a smaller interval set. There's no held-out selection and no estimate of the selection effect. Two specifics. First, BAGEL's optimum is late: Table 10 puts its best GEdit scores at [35,45) (SC 7.94, PQ 6.74, O 6.99), not at [10,16) (7.87, 6.45, 6.81). If the mechanism is "switch where attention has faded and predictions agree", a late optimum needs an explanation — architecture, noise schedule, something. Second, the per-metric optima differ even on 4B (SC [10,16), PQ [15,30), ImgEdit [10,20), CLIP-T [15,30)), so the headline number is one selected point. The selective rule is tuned even more directly: from 111 candidate rules, 8 met a fidelity/preservation target on GEdit, and (0.50, 0.22) was picked. That's a reasonable engineering choice, but the GEdit results for the selective variant are in-sample, and switching 68–69% of edits makes "selective" a bit of a misnomer.
Questions for authors
  1. Where are the quality significance tests promised in Section 4.1? Please report per-metric Wilcoxon/McNemar results with multiplicity control for all three models.
  2. Please run a TOST for judged preservation with a pre-specified margin (say ±0.25/10) per model. With 9B at −0.158 [−0.338, +0.023], can you really claim equivalence? And what is judge–human agreement on the modified preservation prompt?
  3. Under GPT-4.1, E* wins 9B GEdit O (8.10 vs 7.96) and SC (8.79 vs 8.55); under Qwen3.5, E* wins 4B GEdit O (7.80 vs 7.69). What single decision rule makes the interval the right choice, and how should the abstract be worded so it's true in all three settings?
  4. What happens if the interval is chosen per model on a held-out split and evaluated on the others? And why is BAGEL's optimum [35,45) rather than the early-middle window — does that change the mechanism story?
  5. Please rerun the main table with Qwen3.5-35B-A3B captions (0.24 s, equal or better scores) and justify using Gemini (26.76 s) if the cheap captioner is equivalent.
  6. How often does the caption introduce content that isn't in the instruction or the source, as in Figure 9's invented man and Figure 10's "five Asian male security guards"? Please add an unintended-addition rate and an identity/attribute preservation metric.
  7. Is [10,16) an absolute step index? Please test 28 steps and a different sampler, and report the interval in noise-level terms.
  8. Why does "extract" stay at 1.88 (from 1.47) while E* reaches 3.64? For which edit families should a practitioner use the switch?
3 comments
AuthorsOpenReview ↗

Official Response to Reviewer AVtc [1/2]

We thank the reviewer for recognizing the strength of our experimental controls, the image-versus-text decomposition, and the breadth of our evaluation. We address your concerns below and will follow up separately on preservation equivalence and human evaluation (W2/Q2).

W1/Q3. Comparison with E* and the abstract

E* is indeed a strong baseline, as discussed in §4.2. It retains more source content, but T2I intervals achieve higher perceptual quality (PQ), target-caption alignment, and functional correctness across all 3 models (Table 10). These results favor switching when prioritizing these objectives, while E* can be preferable when prioritizing strict source preservation or aggregate scores that depend on it. Fig. 6 isolates the effect of removing the source image, showing larger PQ gains with the same target caption. Thus, switching offers benefits beyond improved instructions. The abstract compares switching with pure editing, against which the fixed interval improves all 6 reported quality scores (Table 2).

Q1. Paired significance tests

Table 2 already shows higher average scores on all 6 quality metrics across all 3 models, supporting the improvements reported in the abstract and contributions. The paired significance tests provide additional evidence about these gains. We omitted their results and now report them below. We compare pure editing with fixed E→T2I→E [10,16), using two-sided Wilcoxon tests for GEdit, MB and ImgEdit, and exact McNemar tests for GIE FC. P-values are Holm-adjusted within each model.

ModelMode / testGEdit SCGEdit PQGEdit OImgEditMB CLIP-TGIE FC (%)
4BPure editing7.6517.5277.1743.7520.3163468.58
Fixed [10,16)7.9977.9997.7343.9620.3209275.35
Adjusted p0.0073<0.001<0.001<0.001<0.001<0.001
9BPure editing8.4707.7467.7784.1370.3186778.35
Fixed [10,16)8.5677.9217.9784.2760.3228281.04
Adjusted p0.4560.03190.0658<0.001<0.0010.0138
BAGELPure editing7.6886.4026.6463.3270.3200366.82
Fixed [10,16)7.8666.4776.8273.4530.3233772.94
Adjusted p0.04940.2670.0326<0.001<0.001<0.001

These results come from newer runs and differ from Table 2 due to variation in judge responses and GPU hardware (4B and BAGEL MB: H100, 9B MB: B200, the rest: A100).

All 6 quality scores improve on all 3 models. After correction, gains are significant for 6/6 metrics on 4B, 4/6 on 9B, and 5/6 on BAGEL (bold: p<0.05).

W3/Q4. Interval choice and BAGEL’s later optimum

Gains occur across multiple intervals; Table 2 uses [10,16) to show broadly consistent improvements with one configuration across all metrics. The full sweeps show how these gains depend on placement and duration. We report these sweeps rather than a held-out evaluation of a selected interval.

We extended Fig. 3 to all 3 models. BAGEL retains substantially more source attention late in denoising than FLUX, while editing and T2I predictions remain broadly aligned. This suggests that a late T2I interval can still remove substantial source influence without strongly redirecting the update, offering an explanation for BAGEL’s stronger GEdit and ImgEdit results with later intervals.

We use [10,16) for selective application to illustrate whether source attention can identify which edits benefit from a given interval. The same approach can be applied to other intervals. Although the 4B thresholds were tuned on GEdit, selective switching also improves all reported quality scores over pure editing on the other 3 benchmarks without further threshold tuning (Table 2).

AuthorsOpenReview ↗

Official Response to Reviewer AVtc [2/2]

Q5. Captioner choice and runtime

We chose Gemini as our default captioner to prioritize caption quality. Appendix D addresses a practical concern: can we retain the gains without increasing inference time? With Qwen captions on FLUX 4B with [10,20), switching improves all 4 benchmarks while running faster than pure editing (Tables 13–14), although fast inference isn’t our central claim. We consider the existing comparison sufficient to address the concern about relying on Gemini.

Q6. Caption errors and preservation

We checked the examples you mention: the man in Fig. 9 (row 4) and the 5 security guards in Fig. 10 (row 4) are already present in the source images, so these captions describe existing content. We mention caption errors in §5 as a potential failure mode, but have not noticed this as a recurring issue. Unintended additions and identity/attribute changes are already covered by our preservation evaluations.

Q7. Step indices and noise levels

Yes, [10,16) denotes absolute, zero-based step indices in our 50-step runs. The table below gives the corresponding noise levels for 50 steps at 1024 × 1024, representative of the pixel counts commonly used across our benchmarks.

ModelEnter T2I: σ10\sigma_{10}Return to editing: σ16\sigma_{16}
FLUX 4B0.96790.9413
FLUX 9B0.96800.9414
BAGEL0.92310.8644

Here, σ\sigma is the noise coefficient, from 1 (pure noise) to 0 (clean). FLUX’s values depend on output resolution; BAGEL’s do not.

The same step interval already spans different noise levels across the evaluated models, with gains under each model’s sampling setup. This supports the benefit beyond one particular noise schedule. Our current experiments keep the total step count fixed, and we leave transfer across step counts and samplers for future work.

Q8. Extraction and practical guidance

ImgEdit’s “extract” category is problematic: most of its editing instructions are ambiguous. E.g. “extract the animal in the image” does not explicitly specify what should remain – the animal or everything else. E* has a strong advantage over the baseline because its improved prompt happens to interpret the instruction in the same way as the VLM judge. E→T2I→E [10,16) has less opportunity to resolve this ambiguity, since it supplies an explicit target for only 6 / 50 steps. This category concentrates an otherwise rare problem – instructions that can imply opposite outputs, so the smaller gain on it does not imply a general weakness on extraction tasks.

For practitioners, switching is useful for edits requiring substantial synthesis and natural integration, whereas source-conditioned editing is preferable when retaining unrelated details is the priority.

AuthorsOpenReview ↗

Follow-up Response to Reviewer AVtc

W2/Q2. Preservation equivalence and human–GPT agreement

Following the suggested margin of ±0.25 on the 0–10 GPT preservation scale, we ran paired TOST equivalence tests comparing pure editing with E→T2I→E [10,16). Each model has 1,079 paired edits, with uncertainty clustered by source image. All 3 models support equivalence in mean judged preservation within this margin after Holm correction across models.

ModelPure editingFixed [10,16)Difference90% CIAdjusted TOST p
FLUX 4B7.75357.6413−0.1121[−0.2254, +0.0011]0.0227
FLUX 9B8.58118.6108+0.0297[−0.0875, +0.1468]0.0030
BAGEL7.72297.8063+0.0834[−0.0332, +0.2000]0.0189

These results come from newer runs and differ from Tables 2 and 16 due to variation in judge responses and GPU hardware (4B and BAGEL: H100, 9B: B200).

Our judge retains GEdit’s preservation criterion and removes only the instruction-fidelity component. To check its agreement with humans, we conducted a blind side-by-side study on 100 randomly selected GIE editing samples for FLUX 4B, with 15 participants and 3 ratings per sample. Participants judged unintended changes and rated them as slight, clear, or large. Human study results

GPT largely agrees with human preferences, especially on large-difference cases, and predominantly assigns equal scores when humans report a tie. Counting both clear and large differences, humans judged that [10,16) preserved unrelated details as well as or better than pure editing on 74% of samples. These judgments complement PSNR, MSE and SSIM, whose sensitivity to pixel-level differences does not directly indicate how much an unintended change matters perceptually.

Our quality evidence also extends beyond GPT-4.1: MagicBrush uses CLIP-T, ImgEdit uses GPT-4o, and the Qwen-judged ablations support the same qualitative conclusions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.