PaperlayerView exact revision on arXiv

Source-linked AI summary

ImageRSI: Verification-Driven Recursive Self-Improvement for Agentic Image Generation

Shuo Han, Zhilin Wang, Peng Ling, Wenjia Gong, Yang Liu, Guanjun Jiang, Yukun Cao, Xike Xie

arXiv:2610.32120v1cs.CVcs.AI

TL;DR

Complex image requests can contain interacting or implicit constraints, and a successful repair does not show that its strategy will transfer. ImageRSI combines verification-driven repair with validated experience reuse and offline judge updates while keeping the generator frozen. It reports 52.6% CSA on MindBench and 65.66% on ImageReq-Bench.

  • Problem

    Complex requests combine explicit and implicit requirements, and neither an individual successful repair nor fixed-prompt final-output evaluation establishes cross-task improvement.

    Source: §1 · p. 1

  • Method

    ImageRSI keeps the image generator frozen and combines five-dimensional verification and requirement-preserving repair, paired-validated experience reuse, and offline judge post-training on accumulated judgment records.

    Source: §1 · pp. 1–2

  • Results

    ImageRSI achieves 52.6% CSA on MindBench with Nano Banana Pro and 65.66% on ImageReq-Bench.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Within the reported settings, CSA increases with experience-library size, and judge post-training further improves end-to-end CSA.

    Source: §5 · p. 9

    The study focuses on improving a given base generator; transferring or adapting experience when replacing it is left as a possible extension.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Complex text-to-image requests remain difficult to satisfy reliably. Verifying and repairing an image can fix an individual output, but a successful repair does not necessarily carry over to other requests. We present ImageRSI, a framework that keeps the image generator frozen and combines three components: five-dimensional verification and repair within each task, an experience library shared across tasks, and offline updating of the judge. Candidate experiences enter the library only after paired A/B validation, and the judge is updated with supervised fine-tuning and GRPO on accumulated image-judgment records and evaluated before reuse. We also introduce ImageReq-Bench, a benchmark of 961 requests built around hard cases in real-world image-generation requests, each paired with an evaluation checklist. ImageRSI achieves a Checklist Strict Accuracy (CSA) of 52.6% on MindBench (with Nano Banana Pro as the generator) and 65.66% on ImageReq-Bench, and we compare system outcomes across judge-training configurations.

1 Introduction

Complex image requests combine interacting and sometimes implicit requirements, while a successful repair does not establish that its strategy transfers to other tasks. ImageRSI combines within-task verification and repair with validated cross-task experience and offline judge updates; the paper reports strong CSA results on both benchmarks.

1 Introduction

  • Complex requests combine interacting explicit and implicit constraints, including factual content, exact text, counts, attributes, and spatial relations.Reliable generation requires inferring intended constraints, checking the image against them, and repairing unmet requirements.Source: §1 · p. 1
  • A successful repair does not establish that its strategy transfers, because stochastic generation, redundant trajectories, and judge errors can distort evidence.The paper identifies validation, applicability conditions, and criteria for revising or retiring experience as necessary safeguards.Source: §1 · p. 1
  • ImageRSI keeps the image generator frozen while coupling five-dimensional verification and targeted repair with cross-task experience and offline judge self-evolution.Repair targets defects while locking parsed requirements; experience admission uses paired A/B validation, and accumulated judgment records support judge post-training.Source: §1 · pp. 1–2
  • ImageReq-Bench contains 961 complex requests, each linked to source records and a checklist for joint requirement satisfaction.The benchmark is built around hard cases in real-world image-generation requests.Source: §1 · p. 2
  • ImageRSI reaches 52.6% CSA on MindBench and 65.66% on ImageReq-Bench.The framework also surpasses the strongest baselines by 10.2 and 9.28 percentage points, respectively.Source: §1 · p. 2

2 Related Work

Prior agentic image-generation work uses planning, tools, iterative evaluation, and sometimes cross-task memory, while fixed-prompt benchmarks assess final-output quality. ImageRSI adds explicit validation and applicability conditions for reusable experience, alongside offline judge post-training.

2 Related Work

  • Agentic approaches address complex image requests through planning, external tools, and iterative evaluation, and some systems reuse memory across tasks.ImageRSI additionally specifies validation and applicability conditions for candidate experience.Source: §2 · p. 2
  • For complex requests, agentic approaches use planning, external tools, and iterative evaluation (Xue et al., 2024; He et al., 2026b; Zhang et al., 2026c); Qwen-Image-Agent reuses memory across tasks (Zhang et al., 2026c), while ImageRSI specifies an explicit validation protocol and applicability conditions for candidate experience.Source: §2 · p. 2
  • Prior work has explored verification, search, trajectory-derived lessons, and iterative improvement of programs or agent context.ImageRSI brings experience-driven improvement to image generation through A/B-tested reusable experience and offline post-training on versioned verification records.Source: §2 · p. 2

3 Method

ImageRSI uses task trajectories to connect frozen-generator image repair, validated experience reuse, and offline judge updates. Its four method stages specify how requirements are preserved, candidate strategies are tested, and later judge versions are trained and redeployed.

3.1 Overview

  • Task trajectories record generation conditions, verification results, and repair actions, supporting both candidate experience and supervision for a later judge version.The experience library updates inference-time context, while judge parameters update offline between collection rounds; the generator stays frozen.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Online Generation and Verification

  • The online workflow parses requirements and holds them fixed across repair rounds, then uses five-dimensional verification to guide defect-targeted prompt repair.Scoring uses the request, parsed requirements, and image evidence; a separate attribution step can inspect the prompt and selected experience to propose repairs. The process stops at acceptance or budget exhaustion and returns the highest-scoring image as fallback if none is accepted.Source: §3.2 · p. 3

3.3 Cross-Task Experience

  • Candidate strategies record triggers, actions, effects, and side effects; paired validation compares images with and without an action under fixed request inputs, generator configuration, and seed.Calibration and a separate holdout account for improvements and regressions before activation; trigger conditions constrain later selection, and outcomes inform revision or retirement.Source: §3.3 · pp. 3–4

3.4 Judger Self-Evolution

  • The judge remains fixed within each task and is updated offline from selected trajectory images paired with score-and-reason annotations.SFT trains on these records, then GRPO optimizes scoring behavior; an evaluated candidate may replace the in-loop judge, while repeated collection–training cycles remain a proposed extension.Source: §3.4 · pp. 4–5

4 Experiments

Experiments evaluate ImageRSI’s benchmark performance, localized repair, experience reuse, and judge post-training. Results show gains over comparison systems and ablations, with performance increasing as the experience library grows and after judge updates.

4.1.1 Experimental Setup

  • The evaluation uses a 431-request ImageReq-Bench test set, while MindBench contains 500 requests; CSA requires every checklist item to pass.ImageReq-Bench has 961 requests paired with checklists, totaling 8,165 check items.Source: §4.1.1 · p. 5

4.1.2 Results on MindBench

  • With Nano Banana Pro, ImageRSI reaches 52.6% CSA on MindBench, 4.4 pp above its Qwen-Image-2.0 version.The Qwen version reaches 48.2% overall CSA, 5.8 pp above the strongest agent baseline, GEMS.Source: §4.1.2 · pp. 5–6, Table 1 · p. 7

4.1.3 Results on ImageReq-Bench

  • ImageRSI reaches 65.66% CSA on ImageReq-Bench, the highest overall result and 9.28 pp above Nano Banana.It also leads GEMS, the strongest agent baseline, by 9.98 pp.Source: §4.1.3 · p. 6, Table 2 · p. 8

4.2 Effectiveness of Verification and Repair

  • Under a matched generation budget, localized repair raises CSA from 53.4% to 65.7% (+12.3 pp), while Best-of-2 resampling gains only 1.6 pp.Generic-feedback repair gains 3.2 pp; removing requirement locking lowers CSA to 57.8%, 7.9 pp below full localized repair.Source: §4.2 · p. 6, Figure 3 · p. 7
  • The two illustrated cases show localized repair fixing initially failed key requirements, while resampling and generic-feedback repair fail to fix them.In case 0656, generic feedback also adds a third assistant and loses two previously satisfied requirements.Source: §4.2 · p. 6, Figure 4 · p. 8

4.3 Experience Reuse

  • CSA rises monotonically from 55.7% to 65.7% as the experience library grows from 0 to 527 entries.The full library improves CSA by 10.0 pp over no experience.Source: §4.3 · p. 7, Figure 3 · p. 7

4.4 Recursive Self-Improvement

  • GRPO reaches 48.2% CSA and 0.403 MAE, improving over SFT100% by 3.4 pp CSA and 0.050 MAE.The trained judge exceeds all evaluated off-the-shelf judges; their CSA ranges from 24.2% to 44.2%.Source: §4.4 · pp. 8–9, Table 3 · p. 8

5 Conclusion

ImageRSI combines within-task verification and targeted repair with validated cross-task experience reuse and offline judge self-evolution, while keeping the image generator frozen. It reaches 52.6% CSA on MindBench with Nano Banana Pro and 65.66% on ImageReq-Bench, though gains from repeated collection–training cycles remain unevaluated.

5 Conclusion

  • ImageRSI couples five-dimensional verification and targeted repair with a validated cross-task experience library and offline judge self-evolution under a frozen image generator.The framework reaches 52.6% CSA on MindBench with Nano Banana Pro and 65.66% on ImageReq-Bench.Source: §5 · p. 9

Limitations

ImageRSI is evaluated with several base image generators, but the stated focus is improving generation quality for a given generator. Reuse across a replacement generator is described as a possible extension, not a contribution of the method.

Limitations

  • The evaluation uses several base image generators.The stated scope is improving generation quality for a given generator.Source: Limitations · p. 9
  • Transfer to a replacement generator is presented as a related application, not as part of the method’s contributions.The authors describe studying which experiences transfer directly and which require adaptation as a possible extension.Source: Limitations · p. 9
  • The proposed extension would reuse existing experience to reduce the cost of accumulating it again for a new generator.Source: Limitations · p. 9

A Implementation Details

The implementation bounds generation and repair, scores images across active dimensions, and admits experience only after paired validation. An optional constraint-aware variant checks requirements individually and repairs only unmet ones.

A.1 Generation and Delivery

  • Each request receives an initial generation and at most two repair rounds; if none is accepted, the highest-scoring judged image is returned as a fallback.Retrieval is limited to four passes per request, and fallback images are not counted as accepted.Source: §A.1 · p. 12

A.2 Scoring and Acceptance

  • Images are accepted at aggregate score 4.5 or higher unless vetoed; low-confidence dimension weights are reduced by 0.7 and normalized over active dimensions.Quality scores of 3 or lower veto acceptance; Alignment scores of 3 or lower also veto unless the parser reports low confidence.Source: §A.2 · p. 12

A.3 Constraint-Aware Verification

  • The optional constraint-aware variant labels each requirement pass, fail, or unknown, then repairs failed or unknown requirements while preserving satisfied clauses.Occluded counts and unobservable relations remain unknown; satisfied requirements are rechecked after repair.Source: §A.3 · p. 12

A.4 Experience Admission

  • Experience candidates need two independent source trajectories and must pass calibration and holdout trials using paired images with shared seeds and upstream inputs.Calibration requires at least three trials and a mean difference of 0.05 or more; holdout requires at least two trials and a nonnegative mean, with no negative differences in either split.Source: §A.4 · p. 12

B Datasets and Evaluation

The paper describes a hard-case image-request benchmark with visually checkable checklists, a strict all-items-pass accuracy measure, and interfaces for selecting and applying experience. Its examples and prompt interfaces specify how strategies and repair feedback are represented in the workflow.

B.1 ImageReq-Bench

  • ImageReq-Bench targets real-world requests whose requirements current generators often fail to satisfy jointly, with each checklist refined into visually checkable items.Checklists contain 2–18 items, and all methods use the same checklist version and fixed test case IDs; category labels are organizational, not a validated difficulty taxonomy.Source: §B.1 · p. 12

B.2 Checklist Strict Accuracy

  • CSA counts a request as passing only when an image exists, every checklist item receives a valid verdict, and all verdicts are true.Malformed evaluator responses are not passes, and the external evaluator is separate from the internal judge; its verdicts are not fed back into that request’s generation.Source: §B.2 · p. 13

C.1 Experience Record Structure

  • Experience selection first uses each entry’s Domain, Task Type, Generation Type, and Failure Pattern, then loads selected entries in full.Source: §C.1 · p. 13

C.2 Example Entries

  • The example experience entries encode triggers, negative triggers, actions, protected constraints, and known side effects for identity anchoring and spatial anchoring.The identity strategy separates descriptions for named subjects; the spatial strategy makes object relations explicit and lists discrete elements when counts matter.Source: §C.2 · p. 13

C.3 Prompt Interfaces

  • The prompt interfaces keep parsed requirements fixed across repair rounds and direct localized prompt changes using image-grounded defect feedback.Dimension scoring uses the request, requirements, and image evidence, while repair suggestions do not change the scores.Source: §C.3 · p. 13
  • Paired experience review scores blinded images generated with and without a candidate using the same seed and upstream inputs.The optional constraint checker returns a status and visible evidence for each atomic requirement; insufficient evidence remains unknown.Source: §C.3 · p. 14
  • The ablation’s generic-feedback repair prompt omits defect information, while experience distillation uses external per-item verdicts on experience-set requests to create reusable advice.Distilled advice is retrieved using the three most similar entries by character-bigram TF-IDF.Source: §C.3 · p. 14

D Judge Training

Judge training combines benchmark experience-split requests with additional requests and pairs selected trajectory images with five-dimensional scores and image-grounded reasons. Deployment supplies confidence values through an inference adapter rather than training on confidence targets.

D.1 Training Data

  • Judge training uses approximately 2,000 requests, combining 530 ImageReq-Bench experience-split requests with about 1,500 additional requests.The added requests draw on WISE taxonomy and prompt design and MindBench knowledge- and reasoning-driven categories.Source: §D.1 · p. 14
  • Selected generation and repair trajectory images receive five dimension-specific integer scores and image-grounded explanations as training targets.Targets cover Alignment, Faithfulness, Quality, Physics, and Aesthetic; images that violate the request remain in training with targets describing the violation.Source: §D.1 · p. 14
  • Training covers scores and reasons only; confidence values are set to constants by an inference adapter when deploying a trained judge.Source: §D.1 · p. 14

D.2 Training Objective

The training objective uses supervised fine-tuning on judge responses, then GRPO to refine scoring behavior. Its reward combines pointwise assessment with conditional within-group ranking while keeping score aggregation and acceptance rules fixed.

D.2 Training Objective

  • SFT minimizes loss on the target response given the judge input and preceding response tokens.Prompt and image tokens are masked from the loss.Source: §D.2 · p. 14
  • GRPO starts from the SFT100% checkpoint, and within-band ranking adjustments apply only to consistent, evidence-supported rankings with sufficient confidence.Score aggregation and acceptance rules remain fixed rather than being learned.Source: §D.2 · pp. 14–15
  • GRPO rewards five-dimensional score accuracy and rationale correctness, while malformed outputs and severe false accepts or critical fabrications or omissions receive lower reward bands.The lower band for severe errors cannot be offset by other dimensions.Source: §D.2 · pp. 14–15

E Experimental Details

The experiments compare baselines, repair strategies, experience-library sizes, and judge aggregation ablations. The supplied results report an increased false-accept rate when the alignment dimension is removed; other passages specify evaluation conditions and table scopes without giving their values.

E.1 Baselines on MindBench

  • Baseline comparisons use the higher of original-paper and reproduced values; ImageRSI’s 12 requests without an image count as failures.Wan2.7-Image’s one unevaluated request also counts as a failure, making its CSA a lower bound; reproduction conditions differ across systems.Source: §E.1 · p. 15

E.2 Repair Strategies

  • Generic-feedback repair asks the rewriter to improve the prompt after a failed check without providing defect information.The second-round strategies share the first-round image and use at most two images per request; the table’s delta is relative to the first-round image only.Source: §E.2 · p. 15

E.3 Experience Library Size

  • The experience-library comparison uses nested prefixes of one fixed random order, with other settings following localized repair.Table 6 reports test-set CSA and delta over using no experience, but the supplied passage gives no values.Source: §E.3 · p. 15, Table 6 · p. 15

F Supplementary Figures

  • 37.1% versus 24.8% false-accept rate when removing alignment rather than using the deployed configuration, in the offline judge ablation.Removing quality or hard vetoes also increases recall and false-accept rate; removing the low-confidence discount changes no decision.Source: Figure 5 · p. 15
  • Figure 5 compares judge configurations by reaggregated dimension scores against external CSA labels.Removing aesthetics and faithfulness are marked as separate nearby configurations.Source: Figure 5 · p. 15, F Supplementary Figures · p. 15

Related papers

Loading 2610.32120v1…