PaperlayerView exact revision on arXiv

Source-linked AI summary

Rubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation

Ritesh Thawkar, Shubham Patle, Shravan Venkatraman, Rao Muhammad Anwer

arXiv:2610.12469v1cs.CV

TL;DR

Image editors still rely on human-edited targets or external reward models for further improvement, although plausible outputs can fail to make the requested edit or preserve content. Rubric-CEPR verifies an editor’s own candidates with structured checks from its internal representations and distills verified samples. It improves ImgEdit performance for Qwen-Image-Edit and Step1X-Edit, with transfer to other editing benchmarks reported for Qwen-Image-Edit.

  • Problem

    Further improvement of instruction-guided image editors relies on costly human-edited pairs or external reward models, which can reward plausible outputs that leave edits undone or alter preserved content.

    Source: §1 · p. 2

  • Method

    Rubric-CEPR uses a Planner, Editor, and fixed internal Critic to check candidate edits against structured edit, old-state-removal, and preservation requirements, then distills verified candidates without human-edited targets or an external training-time reward model.

    Source: §1 · p. 2

  • Results

    ImgEdit improves from 4.36 to 4.60 (+5.5%) for Qwen-Image-Edit-2509, with transfer to GEdit-Bench and Complex-Edit and a +7.8% ImgEdit gain for Step1X-Edit.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The experiments show consistent ImgEdit gains, the largest improvement on object isolation, transfer to GEdit-Bench and Complex-Edit, and a second successful instantiation on Step1X-Edit.

    Source: §7 · p. 10

    The approach requires access to the editor’s multimodal features and latents; evidence covers two multimodal diffusion editors, automatic-judge evaluation, and a single adaptation pass over one source pool.

    Source: §6 · p. 10

Abstract

from arXiv · show

Instruction-guided image editors have become highly capable, yet improving them further still depends on human-edited training pairs or external reward models. Such supervision is costly to obtain and can reward plausible failures: a realistic output may leave the requested change undone or alter content that should be preserved. In this work, we strive to improve a pretrained image editor using only its own generations, without human-edited targets or an external training-time reward model. To this end, we propose a self-evolving framework, named Rubric-CEPR, that verifies the editor's own samples with its internal representations through a rubric-augmented Contrastive Edit-Preservation Reward (CEPR). A Planner proposes structured edit instructions from unlabeled images, the Editor samples multiple candidate edits, and a frozen Critic scores each candidate with decomposed rubric checks for edit realization, removal of the old state, and content preservation, using features already exposed by the editor. Non-compensatory gates reject infeasible candidates, and the best verified candidate is distilled into the editor through lightweight adapter training. On Qwen-Image-Edit, Rubric-CEPR improves ImgEdit from 4.36 to 4.60 (+5.5%), with a +24.9% gain on object isolation, and transfers to GEdit-Bench and Complex-Edit. The same procedure also improves Step1X-Edit by +7.8% on ImgEdit. We hope our approach will serve as a solid baseline for image editors that improve themselves from their own verified samples. Our code is publicly available at $\href{https://riteshthawkar.github.io/Rubric-CEPR/}{\text{this URL}}$

1 Introduction

Image editors must make requested changes while preserving untouched content, but current improvement relies on costly external supervision that can accept plausible failures. Rubric-CEPR instead verifies the editor’s own candidates with structured checks before distilling them, improving benchmark performance across two editors.

1 Introduction

  • Instruction-guided editors must perform the requested change while preserving content the instruction leaves unchanged, yet realistic outputs can leave the edit undone.Curated source–edit pairs and learned reward models or multimodal judges are costly to scale, and scalar plausibility or relevance can favor unchanged, partial, or content-discarding outputs.Source: §1 · p. 2
  • ImgEdit improves from 4.36 to 4.60 (+5.5%) for Qwen-Image-Edit-2509, with a +24.9% gain on object isolation; gains also transfer to GEdit-Bench and Complex-Edit.Applying the procedure to Step1X-Edit improves ImgEdit by +7.8%.Source: §1 · p. 2
  • Rubric-CEPR uses a Planner, Editor, and fixed internal Critic in a propose–edit–verify–distill loop, without human-edited targets or an external training-time reward model.The best verified candidate supervises a lightweight editor adapter through the editor’s native denoising objective.Source: §1 · p. 2
  • Its decomposed rubric checks edit realization, removal of the old state, and preservation separately, with non-compensatory gates that reject candidates failing an applicable check.The Critic uses editor-exposed Qwen-VL features and VAE latents.Source: §1 · p. 2
  • Training the Planner increases the rate of feasible targets by 62.4%.The evaluation reports this result alongside gains on ImgEdit, GEdit-Bench, and Complex-Edit.Source: §1 · p. 2

2 Related Work

Prior image-editing post-training uses paired data, preference or reward supervision, or judges that may reward plausible but incorrect edits. Rubric-CEPR differs by verifying single-turn edits against structured specifications with a fixed internal Critic.

Instruction-guided image editing

  • Instruction-following editors use natural-language commands and multimodal backbones, but pretrained editors lack a mechanism to keep improving from their own outputs without new paired data.Such editors are commonly evaluated on ImgEdit, GEdit-Bench, and Complex-Edit.Source: §2 · p. 2

Reward models and editing post-training

  • Editing post-training methods use expert preference annotations, separate reward models, external multimodal judges, or preference and reward optimization, while a scalar score can still favor plausible outputs that miss the edit.Examples include EditReward [41], EditScore [28], UniWorld’s Edit-R1 framework [25], ImageEdit-R1 [23], and NP-Edit [52].Source: §2 · p. 3

Self-training and preference optimization

  • Existing self-training and preference methods address language, perception, generation, tool-based retouching, or multi-turn editing, rather than verifying single-turn edits against structured specifications.For single-turn editing, the reward must distinguish a correct edit from a plausible but unedited output.Source: §2 · p. 3

Our Approach

  • Rubric-CEPR uses unlabeled source images and a Planner–Editor–Critic loop; its fixed Critic checks edit realization, old-state removal, and preservation separately, and gates reject candidates that fail an applicable check.Verified targets are distilled with the editor’s native denoising objective.Source: §2 · p. 3

3 Method

Rubric-CEPR improves a pretrained editor by generating and verifying its own candidate edits from unlabeled images, then distilling verified targets. Its Planner and Editor use lightweight LoRA adapters, while the reference features and Critic remain frozen.

3 Method

  • The framework uses unlabeled source images to generate training targets from the pretrained editor’s own verified samples, without human-edited targets or external reward models.A Planner proposes instructions and structured specifications, an Editor samples candidate edits, and a fixed Critic returns rewards and feasibility indicators.Source: §3 · p. 3

3.1 Structured Proposals and Candidate Selection

  • The Planner’s structured specification records the edit type, entities, target region, required and forbidden after-states, and content to preserve.These fields determine the Critic’s atomic prompts rather than serving as a second edited target; the primary run uses K = 4 candidates and contributes no Editor target if none pass.Source: §3.1 · p. 4

3.2 Frozen Internal Features

  • The Critic uses frozen Qwen-VL features for semantic support and gain, plus VAE latents for preservation and locality checks.The primary object-isolation reward also uses programmatic background-purity and object-completeness checks; the broader verifier-bank variant additionally uses editor-side VLM token log-probabilities and an open-vocabulary detector for addition analysis.Source: §3.2 · pp. 4–5

3.3 Decomposed Rubric and Feasibility Gates

  • The rubric checks required and forbidden states, counterfactual edit prompts, and semantic and latent preservation; every applicable feasibility check must pass.The latent-change mask estimates changed regions rather than providing a ground-truth edit region, and failed gates make candidates ineligible regardless of visual plausibility.Source: §3.3 · p. 5

3.4 Editor and Planner Updates

  • The Editor receives weighted denoising SFT on selected targets, while the Planner receives reward-weighted token SFT on filtered proposal traces.Rejected candidates do not enter the Editor loss; the Planner’s band-pass set excludes proposals for which every candidate fails or every candidate passes.Source: §3.4 · p. 5
  • 4.36 to 4.60 (+5.5%) on ImgEdit; the largest gain is on extraction, from 3.41 to 4.26 (+24.9%).The overall result is +0.24 ± 0.02 SD over three seeds; extraction is the weakest base family.Source: Table 1 · p. 6

4 Experimental Setup

Training uses unlabeled, off-benchmark source images, with the Editor’s own Critic-selected candidates as targets. Evaluation uses benchmark-specific metrics and protocols, with matched generation settings for the base and adapted editors.

4.1 Training Data and Implementation Details

  • Training uses unlabeled, off-benchmark images and no human-edited targets; the primary Qwen-Image-Edit-2509 run uses K = 4, 400 Editor SFT steps, and 16 Planner steps.A second instantiation applies the procedure to Step1X-Edit with its own VLM and VAE, without importing Qwen-Image-Edit weights.Source: §4.1 · p. 6

4.2 Evaluation

  • Evaluation uses one generation per instruction with matched settings; ImgEdit uses GPT-4o ratings, GEdit-Bench uses GPT-4.1 VIEScore across 11 tasks, and Complex-Edit uses compound edits and 0–10 metrics.Step1X-Edit is reported on GEdit-Bench and ImgEdit; results from other papers retain their original protocols and are not rebased.Source: §4.2 · p. 6

4.3 Baselines and Controls

  • Comparisons include the pretrained base under the same protocol, published image editors, and UniWorld, which post-trains the same base with an external multimodal reward.Controls include ungated self-training, a 210-pair candidate-ranking probe, a Planner-training control, and a blinded 64-source proposal-grounding audit.Source: §4.3 · p. 6

5 Results and Analysis

Rubric-CEPR improves Qwen-Image-Edit on ImgEdit and transfers to other benchmarks and Step1X-Edit. The analyses find that verified target selection, best-of-K distillation, and reward coverage shape the gains.

5.1 Main Results

  • 7.39 to 8.31 (+12.4%) on GEdit-Bench and 8.77 to 8.97 (+2.3%) on Complex-Edit, with every reported metric improving without further adaptation.Complex-Edit's largest gains are perceptual quality (+4.9%) and identity preservation (+1.8%).Source: Table 2 · p. 7, §5.1 · p. 7
  • 4.36 to 4.60 (+5.5%): Rubric-CEPR improves every ImgEdit family, with extraction rising from 3.41 to 4.26 (+24.9%).Across three seeds, the overall score is 4.58, 4.60, and 4.62; gains close 22%–53% of each family's remaining gap to the 5.0 ceiling.Source: §5.1 · p. 7
  • 6.69 to 7.24 (+8.2%) on GEdit-Bench and 3.86 to 4.16 (+7.8%) on ImgEdit show gains when the procedure is applied to Step1X-Edit.The procedure uses Step1X-Edit's own VLM and VAE features, without Qwen-Image-Edit weights.Source: §5.1 · pp. 7–8, Table 3 · p. 8

5.2 Why Verified Targets Matter

  • On a 210-pair probe, the verifier-bank gate accepts 50.5% of valid edits and none of the invalid controls, while naive self-training gains less than Rubric-CEPR.The reward selects the stronger edit in 84.8% of pairs versus 36.7% for a naive product; removing the editor-side VLM check admits 4.8% of corrupted candidates.Source: §5.2 · p. 8, Table 4 · p. 8
  • Reward-best extraction scores rise from 3.41 at K=1 to 4.45 at K=8, while single-shot distillation reaches 4.26, on par with best-of-4.The distilled result is a +24.9% gain and requires no inference-time sampling.Source: §5.2 · p. 8, Figure 3 · p. 9

5.3 Proposal Learning and Reward Coverage

  • A four-type verifier bank retains the isolation improvement (+25.8%) but has a smaller composition gain (+3.2% versus +8.4%) and similar overall gain (+5.3% versus +5.5%).Broader reward coverage does not compound the benefit of the aligned reward.Source: §5.3 · p. 9

6 Limitations

Rubric-CEPR can amplify only successes already present in the editor’s sampling distribution, and gains depend on whether internal signals can verify success. The evidence is bounded by the available features, evaluated editors, judges, and training setup.

6 Limitations

  • Sampling more candidates cannot recover an edit absent from the editor’s sampling distribution.Under independent draws, the chance of finding a valid target is 1 − (1 − p)^K, which is zero when p = 0.Source: §6 · p. 9
  • On adjustment edits, tested internal signals recover at most 30% of oracle headroom.The approach is most effective where success is internally checkable.Source: §6 · p. 9
  • The approach requires access to editor features and latents; evidence covers two diffusion editors, automatic judges, and one adaptation pass over one source pool.The authors identify independent human assessment, matched reward controls, and multi-round learning curves as further evidence that would strengthen claims.Source: §6 · p. 10

7 Conclusion

Rubric-CEPR improves pretrained editors by distilling their own verified successes without human-edited targets or an external training-time reward model. The appendix outlines diagnostics of sampling, target construction, and the self-evolution pipeline.

7 Conclusion

  • Rubric-CEPR couples a Planner, Editor, and fixed internal Critic in a gated propose–edit–verify–distill loop.The conclusion reports gains on ImgEdit, transfer to GEdit-Bench and Complex-Edit, and a second successful instantiation on Step1X-Edit.Source: §7 · p. 10

A Supplementary Experiments and Details

  • The appendix covers sampling and target-selection diagnostics, reward-signal and headroom studies, algorithms, evaluation protocols, implementation details, ablations, and qualitative results.It distinguishes the primary internal-only model from broader verifier-bank analyses.Source: A Supplementary Experiments and Details · p. 14

B Sampling and Target Construction

  • The sampling and target-construction analysis examines recoverable successful edits, single-shot internalization, and whether reward gates reject invalid pseudo-labels.It frames these as properties of the pipeline under study.Source: B Sampling and Target Construction · p. 14

B.1 Naive-Loop Target Distribution

The supplementary studies examine naive target quality, sampling headroom, and the internal reward’s ability to select usable distillation targets. They show that headroom and selection quality vary across edit families.

B.1 Naive-Loop Target Distribution

  • The naive four-round loop ends with 546 rejected candidates versus 302 accepted records, and clean removal and replacement targets are scarce.The figure reports only 3 and 2 clean targets for removal and replacement, with 92% and 96% rejected, respectively; color edits dominate the accepted pool.Source: Figure A1 · p. 14, §B.1 · p. 14

B.2 Best-of-K Sampling Diagnostics

  • On full ImgEdit extraction, reward-best selection rises from 3.41 at K=1 to 4.45 at K=8, while the distilled checkpoint scores 4.26 with one generation.The comparison motivates reward-verified distillation over best-of-K inference, which requires multiple samples and an inference-time selector.Source: §B.2 · p. 14

C Extended Reward-Signal and Headroom Studies

  • The extended studies probe family-level headroom, reward-based selection, adjustment self-verifiability, and whether the headroom pattern extends to Step1X-Edit.The section frames these as diagnostics supporting the main paper’s evidence.Source: C Extended Reward-Signal and Headroom… · p. 15

C.1 Per-Family Recoverable Headroom

  • Object isolation exposes +1.86 best-of-5 headroom, while near-ceiling families gain only +0.25 to +0.28.The six-family probe covers 525 of 737 benchmark examples; its weighted best-of-5 selection projection is +0.51, assuming uncovered families remain unchanged.Source: §C.1 · p. 15

C.2 Reward-Verified Target Selection

  • 84.8% pooled selection accuracy, versus 50% chance, is achieved by the verifier-bank reward; a naive product of reward components selects correctly only 36.7%.The verifier-bank reward reaches 80–97% on six families and 60% on removal; the naive product falls below chance on near-ceiling families by over-weighting preservation.Source: §C.2 · p. 15

C.3 The Self-Verifiability Boundary

The studies show that reward quality and internal checkability bound how much sampling headroom can be distilled. The appendix also specifies a gated training loop and separates evaluation protocols across benchmarks.

C.3 The Self-Verifiability Boundary

  • On adjustment edits, the best internal signal recovers only 30% of oracle headroom, while semantic CEPR signals can rank perceptually worse samples higher.The authors caution that this probe does not establish that any edit family is universally unverifiable; explicit after-state criteria are available for extraction, removal, and addition.Source: §C.3 · pp. 15–16, Table A2 · p. 16

C.4 Cross-Backbone Diagnostic: Step1X-Edit

  • Object isolation is the weakest family on both backbones—1.64 on Step1X-Edit versus 2.97 on Qwen-Image-Edit—and removal is second-weakest on both.On Step1X-Edit, removal’s reward correlation is 0.64 and self-distillation recovers 57% of oracle headroom; object isolation’s correlation is 0.16 and its distilled gain is smaller.Source: §C.4 · p. 16

D Algorithms

  • The round-based algorithm samples instructions and candidate edits, scores them with a fixed internal Critic, selects the highest-ranked gate-passing candidate, and trains adapters on accepted records plus replay or identity records.The gated reward aggregates supported edit and preservation checks and rejects candidates if any required check fails.Source: D Algorithms · pp. 16–17

E Evaluation Protocols

  • ImgEdit, Complex-Edit, and GEdit-Bench use VIEScore, but results are interpreted only within their respective benchmark and scorer protocols.The 11-task GEdit-Bench study uses a different judge from ImgEdit, while Complex-Edit evaluates compound edits, instruction following, identity, and perceptual quality.Source: E Evaluation Protocols · p. 17

E.1 Self-Training Procedures and Signal Sources

Rubric-CEPR outperforms ungated self-training and generated-target DPO under the reported protocols. It separates training-time signals from evaluation judges, uses a documented adapter-and-data pipeline, and gates candidates before they can become training targets.

E.1 Self-Training Procedures and Signal Sources

  • +5.5% on ImgEdit and +12.4% on GEdit-Bench are Rubric-CEPR’s gains over the Qwen-Image-Edit base, larger than the naive loop’s +2.0% and +10.5%; DPO is roughly neutral at +0.7%.The naive-loop GEdit-Bench evaluation uses a different scorer, and the DPO result uses a smaller ImgEdit diagnostic split.Source: Table A4 · p. 18
  • Target selection uses editor-exposed features, VAE latents, and image statistics, while external judges are reserved for benchmark evaluation.The signal table distinguishes training-time reward signals, analysis signals, and final evaluation judges.Source: §E.1 · p. 17, Table A5 · p. 18

F Implementation Details

  • The reported setup uses Qwen-Image-Edit-2509, four candidates per proposal, and an attention-only rank-16, alpha-16 LoRA anchored to the warm-start checkpoint.The self-distillation manifest records each accepted target’s source, instruction, specification, reward components, gate decisions, and replay flag; rejected candidates stay separate from SFT.Source: F Implementation Details · p. 18

G.1 Reward Gate Components

  • The verifier-bank gate accepts 50.5% of valid edits while accepting none of the three negative controls; accepted candidates become subsequent denoising targets.The procedure prioritizes zero negative-control acceptance over looser valid-edit recall.Source: §G.1 · p. 18

G.2 Per-Type Self-Distillation

Self-distillation gains track where the base editor has headroom, while broader reward coverage redistributes rather than compounds gains. The proposal audit measures grounding and coverage, and GEdit-Bench gains vary by task without regressions.

G.2 Per-Type Self-Distillation

  • Yes/no token log-probabilities reduce candidate-pair ties to 7% and correlate 0.32 with the evaluation judge, versus 93% ties and 0.06 for verbalized confidence.These are descriptive measurements of the broader verifier-bank signal, not an independent human-alignment benchmark.Source: §G.2 · p. 19
  • +5.5% overall and +24.9% on extraction are reported for the aligned object-isolation reward; the broader four-type reward yields +5.3% overall and +25.8% on extraction, but composition gains fall from +8.4% to +3.2%.The broader reward redistributes smaller gains across secondary families rather than increasing the aggregate gain.Source: §G.2 · p. 18, Table A8 · p. 20
  • Outside object isolation, the primary checkpoint gains +0.13 to +0.34 across the other eight families, while the 42-item off-target probe changes by 0.11.Across nine families, base score and gain have descriptive correlation r = −0.94, but the small family count and unequal headroom preclude a universal prediction rule.Source: §G.2 · p. 20

G.3 Planner Proposal Audit

  • The blinded 64-source audit finds correct grounding for 58/64 sources (κ = 0.82), with 63 valid emissions, 61 complete specifications, and 60 unique proposals.All 13 scheduled proposal types occur; the audit measures grounding and coverage rather than human preferences over final edited images.Source: §G.3 · p. 20

G.4 Per-Task GEdit Transfer

  • Across 11 GEdit-Bench tasks, gains are strongest for style change, human photo editing, and tone transfer, weakest for color alteration, and no task regresses.These task-level results are reported in Table A9; the strongest gains are among tasks with lower base scores.Source: §G.4 · p. 20, Table A9 · p. 20

H Additional Qualitative Samples

The qualitative examples show complementary outcomes: Rubric-CEPR completes high-headroom edits that the matched baseline misses and retains competence on edit families where the baseline is often already strong.

H Additional Qualitative Samples

  • On unused ImgEdit examples, Rubric-CEPR completes requested extraction, removal, and background edits when the baseline fails to realize the operation.The extraction examples use the object-isolation checkpoint; removal and background examples use the final checkpoint.Source: H Additional Qualitative Samples · p. 20, Figure A2 · p. 21
  • On replacement, addition, adjustment, action, and style-transfer examples, self-distillation preserves baseline competence while improving placement, localization, or visual faithfulness.These examples cover families where the baseline is often already strong.Source: H Additional Qualitative Samples · p. 20, Figure A3 · p. 21
Loading 2610.12469v1…