acceptodds
Under review as a conference paper at ICLR 2027

When Reflection Fails: Improving Self-Reflective Unified Multimodal Models with Correction Amplification and Fallback

Abstract

Self-reflective unified multimodal models (UMMs) offer a promising paradigm for complex image generation by enabling a single model to inspect its own outputs and iteratively refine them. However, they often struggle with generation tasks involving multiple objects and spatial relationships. To mitigate this problem, existing approaches typically rely on utilizing agents to perform reflection and generation separately, rather than exploiting the unified model's intrinsic capability. We revisit the source of these failures within the unified model and identify two distinct bottlenecks: reflection failure, where the model fails to diagnose errors in its reflection, and correction failure, where the model cannot effectively translate correction feedback into visual outputs. Based on this decomposition, we propose RECAF, a training-free framework that enhances both stages while retaining the unified model as the central component for reflection and generation. Experiments on GenEval, GenEval++, and GenEval2 demonstrate that RECAF consistently improves self-reflective image generation, outperforming baselines by 5.0, 5.7, and 3.5 points, respectively.

Primary area:
generative models
Keywords:
Unified Multimodal Model, Self-Reflection, Self-Reflective Unified Model, Compositionality

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_x8WAOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates two failure modes in self-reflective unified multimodal models for compositional image generation: failing to identify errors in generated images and failing to execute corrections after identifying those errors.

To address these issues, the authors propose RECAF, a training-free framework that employs external verification to assess generated images and scales the KV cache states associated with correction instructions to strengthen their influence on subsequent generation.

Experiments on GenEval, GenEval++, and GenEval2 demonstrate consistent performance improvements, with ablation studies supporting the effectiveness of both components.

Critical strengths
  1. Clear problem analysis. The paper distinguishes between reflection failures and correction failures and provides empirical evidence for both. This offers a useful perspective on the limitations of unified multimodal models in iterative image generation.

  2. Effective training-free approach. RECAF achieves substantial improvements across multiple compositional image generation benchmarks without updating model parameters. The ablation studies also demonstrate the complementary benefits of its two components.

Critical weaknesses
  1. Limited methodological novelty. The two core components essentially combine external model-based verification with KV cache scaling, both of which build upon existing techniques. The distinction between reflection failures and correction failures is also relatively intuitive. Despite the promising experimental improvements, the paper does not sufficiently demonstrate that combining these components constitutes a substantial methodological contribution rather than an effective engineering integration. Furthermore, the lack of comparisons with simpler correction enhancement strategies weakens the justification for the proposed design.

  2. Insufficient evaluation of generalization. The experiments primarily focus on compositional constraints involving object counts, colors, and spatial relationships, without evaluating more complex semantic reasoning tasks. More importantly, the fallback mechanism can rely on benchmark metadata to obtain the constraints to be verified, whereas real-world user prompts typically do not provide such structured information. Therefore, the applicability of the proposed method to open-ended generation scenarios remains unclear.

Questions for authors
  1. What is the key technical novelty of RECAF compared with existing external verification and KV steering methods? Could the authors compare their approach with simpler alternatives, such as directly strengthening correction instructions or modifying attention logits, to justify the necessity of the proposed design?

  2. How does RECAF automatically extract verifiable constraints from natural language prompts when benchmark metadata is unavailable? Can the method maintain comparable performance improvements under this setting?

  3. Have the authors evaluated RECAF on WISE [1], T2I-ReasonBench [2], or other benchmarks involving complex semantic understanding and reasoning? Such evaluations would help determine whether the proposed method generalizes beyond relatively simple compositional constraint checking.

[1] WISE: World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation

[2] T2I-ReasonBench: Benchmarking Reasoning-Informed Text-to-Image Generation

5 comments
AuthorsOpenReview ↗

Reply to Reviewer x8WA - Part 1

We thank the reviewer for their thoughtful and constructive feedback. We particularly appreciate the reviewer’s recognition of our work with “clear problem analysis and offers a useful perspective on the limitations of unified multimodal models in iterative image generation’’ and acknowledging RECAF as an “effective training-free approach’’ that “achieves substantial improvements across multiple compositional image generation benchmarks without updating model parameters’’. We address the reviewer’s concerns in detail below.

W1&Q1:

We would like to clarify that our key methodological contribution is not the individual use of external verification or KV-cache scaling, but a failure-aware correction framework that identifies and addresses two distinct bottlenecks in iterative generation.

Specifically, our contribution lies in empirically characterizing these failure modes and developing targeted interventions to address them. As shown in Secs. 3.1 and 3.3, successful error identification does not necessarily lead to successful correction, and correction failures can persist even when attention is concentrated on the target region. This reveals an important limitation of approaches that rely solely on improved error detection or spatial attention and remain underexplored.

Motivated by these observations, our framework addresses the two failure modes through complementary mechanisms: external verification provides explicit feedback on unmet requirements, while KV-cache scaling aims to strengthen the influence of corrective instructions during generation. The key insight is that improving error detection and improving correction execution are distinct objectives that require different interventions. Our method systematically connects these interventions.

We therefore emphasize that our contribution is a failure-driven methodological design, rather than a claim of novelty for its individual building blocks. The experimental improvements demonstrate the practical value of this design.

Comparison with simpler correction-enhancement strategies. To further justify our design, we conduct additional experiments with three simpler alternatives in GenEval and GenEval++:

  • Instruction emphasis: Strengthen the original correction instruction by adding the phrase "Pay attention to the instruction!" at the beginning.

  • Instruction repetition: Repeat the original correction instruction three times to increase its textual emphasis.

  • Attention-logit bias: Add a positive bias of ln(2) to the attention logits corresponding to correction-instruction tokens, rather than scaling their key and value representations.

Results on GenEval

SettingSingle Obj.Two Obj.CountingColorsPositionColor Attri.Overall
UniCoT99.092.083.088.064.067.082.0
Instruction emphasis99.095.078.093.064.065.082.0
Instruction repetition99.095.078.093.062.065.081.0
Attention-logit bias99.095.079.093.063.065.082.0
RECAF99.092.080.094.089.078.089.0

Results on GenEval++

SettingCol.Cnt.Col./Cnt.Col./Pos.Pos./Cnt.Pos./SizeMulti-Cnt.Overall
UniCoT57.5555557.5505062.555.4
Instruction emphasis60.060.060.057.557.547.557.557.1
Instruction repetition57.557.557.560.060.050.060.057.5
Attention-logit bias57.557.552.557.557.552.557.556.1
RECAF60.065.065.065.057.547.567.561.1

The results above show that RECAF consistently achieves higher overall accuracy than all three alternatives. The results suggest that simply emphasizing or repeating correction instructions, or increasing their attention logits, is insufficient to reproduce the improvements achieved by joint K/V amplification in the settings we evaluated. We have incorporated these comparisons into Sec. 4.3 and App. D. To demonstrate RECAF offers better performance under matched generation budgets, we have also conducted experiments to compare verifier-guided resampling and Best-of-N, see Reply to Reviewer WRRa W1&Q2.

AuthorsOpenReview ↗

Reply to Reviewer x8WA - Part 2

Q2:

To clarify, we do not use official benchmark evaluator components or outputs. We used the metadata to construct the requirement sets for the verifier. However, we found that such metadata is not necessary. The same requirement sets can be extracted directly from prompts using a rule-based parser, as in [1].

To demonstrate this, we employ a rule-based parser [1]. To evaluate the accuracy of the parser-extracted requirement set against a benchmark metadata-derived requirement set, we follow [2] and calculate the requirement-set exact-match accuracy. The results from all three datasets show that the parser-derived requirement set achieves 100% exact-match accuracy against the metadata requirements, removing the need for the metadata from benchmarks. We revised our paper accordingly, see App. B.2.

[1] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating Semantically Precise Scene Graphs from Textual Descriptions for Improved Image Retrieval. In Proceedings of the Fourth Workshop on Vision and Language, Portugal. Association for Computational Linguistics. 2015

[2] Anderson, Peter, et al. Spice: Semantic propositional image caption evaluation. European conference on computer vision. 2016

W2&Q3:

We thank the reviewer for this suggestion. We would like to clarify that RECAF primarily targets reflection and correction failures in compositional image generation (as stated in the initial submission), rather than improving the underlying UMM's knowledge or reasoning capabilities. Benchmarks such as WISE and T2I-ReasonBench introduce a different challenge: generation prompts may require the model to infer appropriate visual content from implicit world knowledge or abstract semantic descriptions, rather than satisfy explicitly stated object, attribute, and spatial constraints.

Our original SAM-based verifier is designed for directly observable compositional constraints and is therefore not suitable for evaluating such knowledge-intensive requirements. However, the fallback mechanism itself is verifier-agnostic: its role is to identify violations that the UMM's intrinsic reflection misses, and the external verification component can be adapted to the task at hand.

To examine this adaptability, we conduct additional experiments on WISE and T2I-ReasonBench, replacing the original SAM-based verifier with a VLM-based semantic verifier, Qwen3.5 35B. We follow the official repo for evaluation. For correction amplification we use α=1.01\alpha = 1.01 without any hyperparameter search. We evaluate WISE and T2I-ReasonBench across three different random seeds to assess the stability of our results.

The results in the following tables show that RECAF improves the overall WISE score from 71.2 to 73.1, and the overall T2I-ReasonBench score from 52.5 to 54.9. These results suggest that adapting the fallback to semantic verification can help identify and correct errors that the model's intrinsic reflection misses, even in knowledge-intensive generation settings. We have incorporated the results as a generalization to reasoning image generation, along with the experimental configuration in App. F of the revised version.

WISE Results

SettingOverallCulturalTimeSpaceBiologyPhysicsChemistry
Uni-CoT71.2 ± 0.2572.9 ±0.3268.2 ±1.374.8 ±1.467.9 ±1.773.3 ±0.365.4 ±1.7
RECAF73.1± 0.2573.6± 0.5567.7± 0.178.3± 0.771.8± 1.7579.1± 1.5368.0± 1.56

T2I-ReasonBench Results

SettingOverallIdiomTextualEntityScientific
Uni-CoT52.5 ± 0.5145.4±1.3749.8±1.0752.5±0.8562.1±1.87
RECAF54.9±1.0548.1±0.9452.8±0.2354±1.8764.6±1.16

Dear Reviewer, as the discussion period is coming to an end, we would appreciate hearing from you. Thank you!

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

This paper investigates the reliability of self-reflective unified multimodal models in compositional image generation. By analyzing Uni-CoT's iterative generation process, the authors distinguish generation failures into reflection failure and correction failure, and propose a training-free framework, RECAF. Specifically, they use an external visual verifier to check compositional constraints to reduce erroneous termination, and scale the KV Cache associated with correction instructions to strengthen the influence of correction information on subsequent generation. Experiments show that the method achieves certain performance improvements on GenEval, GenEval++, and GenEval2.

Critical strengths
  1. The paper clearly decomposes failure modes in self-reflective image generation. Distinguishing "reflection failure" from "correction failure" has diagnostic value, helping identify which stage of iterative generation breaks down rather than attributing all failures vaguely to insufficient self-reflection. This improves problem definition and analytical clarity.
  2. The method requires no additional training and achieves stable gains on multiple compositional image generation benchmarks. RECAF outperforms compared baselines on GenEval, GenEval++, and GenEval2, indicating the practical effectiveness of its inference-time interventions and its potential applicability to existing self-reflective UMMs.
Critical weaknesses
  1. The mechanistic justification for Correction Amplification is still insufficient. The paper infers that correction failures are mainly caused by weak corrective signals from the observation that attention already focuses on the relevant region. However, this does not rule out other causes, such as limited spatial editing capability or interference from the original visual condition. The current ablations show that KV amplification is empirically effective, but do not directly validate the proposed failure mechanism.
  2. The independent contribution of the proposed design over simpler alternatives remains unclear. Although the paper provides component-wise ablations, it does not compare against simpler controls, such as directly using verifier-detected errors as textual feedback without KV amplification. Since Fallback may also introduce additional refinement rounds, matched-budget comparisons would help determine whether the gains come from the proposed intervention itself or from stronger feedback and extra correction opportunities.
Questions for authors
  1. What more direct evidence supports the claim that Correction Failure is mainly caused by insufficient corrective-signal strength, rather than limited spatial editing capability or interference from the original visual condition? Does KV Amplification actually increase the effective influence of correction instructions during regeneration?
  2. When the external verifier has already identified a specific error, how does a simpler baseline perform if the detected error is used directly as textual feedback with the original regeneration process, without KV Amplification? Can such a baseline achieve performance comparable to RECAF?
  3. How much additional inference overhead does RECAF introduce compared with the base model, in terms of average refinement rounds, external model calls, and end-to-end latency? Do the reported gains remain under a matched generation or inference budget?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer Gbub - Part 1

We thank the reviewer for the time invested in the review and suggestions. We appreciate the reviewer’s recognition that our analysis of failure modes has “diagnostic value, helping identify which stage of iterative generation breaks down rather than attributing all failures vaguely to insufficient self-reflection. This improves problem definition and analytical clarity.” and their positive evaluation of the method’s “practical effectiveness of its inference-time interventions and its potential applicability to existing self-reflective UMMs”. We address the reviewer’s concerns in detail below.

W1&Q1:

We agree with the reviewer and would like to clarify that our paper does not claim that insufficient corrective signal strength is the main cause of failure. Rather, as discussed in Sec. 3.1 and Sec. 3.3, we observe that correction failures can occur even when attention is already focused on the target region. This observation motivates our hypothesis that insufficient corrective influence may be one contributing factor, without ruling out other factors such as limited spatial editing capability or interference from the original visual condition.

To provide further evidence that KV amplification of the instruction influences the execution of correction instructions, we examine whether scaling correction-related KV representations improves the actual execution of the editing instructions on GenEval, following the correction-fidelity evaluation in Sec. 3.1.

As shown in the following table, amplification increases fully followed instructions from 32% to 51% and reduces the not followed ones from 59% to 32%, demonstrating that correction amplification effectively improves instruction following. However, still 32% remain unfulfilled even with amplification, indicating that strengthening corrective signals alone cannot resolve all correction failures. This observation is consistent with the reviewer's suggestion that other factors, such as limited spatial editing capability or interference from the original visual condition, may also contribute to unsuccessful corrections.

Correction FidelityNo AmplificationAmplification
Followed32%51%
Partially followed9%17%
Not followed59%32%

Overall, these results support our hypothesis that strengthening correction-related representations can improve the realization of the correction instructions.

W2&Q2:

We would like to clarify that the suggested baseline (detected error is used directly as textual feedback) has been present in our ablation study (Fig. 7(a)), under the setting “Reflection w/ Fallback.” Specifically, when the external verifier identifies a violation, the detected error is converted directly into a textual editing instruction and passed to the original Uni-CoT regeneration process without KV Amplification.

AuthorsOpenReview ↗

Reply to Reviewer Gbub - Part 2

W2&Q3:

Compare with verifier-guided resampling/selection: To disentangle the benefits of additional generation attempts, we perform comparisons against verifier-guided resampling and Best-of-N selection.

Verifier-guided resampling: Each generated image is evaluated using the same verifier employed in the Fallback component. If violations are detected, the model generates a new image from scratch using a different random seed. This process repeats until an image satisfies all verifiable constraints or the maximum generation budget is reached. If no candidate passes, we select the one with the fewest detected violations.

Best-of-N selection: The model independently generates N candidate images using different random seeds. The candidate images are evaluated using the same verifier that we used in the Fallback component, and the image with the highest verifier-derived constraint satisfaction is selected as the final output. We choose N to match the maximum number of image generations (1 initial+ 5 rounds) we set in RECAF, or set the budget to match RECAF (average 2 for GenEval, 3 for GenEval++, 5 for GenEval2).

As an additional ablation, we evaluate Always Fallback + Amplification, where the model always applies the fallback and amplification. We refer to it as RECAF(always Fallback + Amplification) below.

We compare across three datasets and report overall accuracy, the average number of generated images per prompt, and the average number of verifier calls (SAM+BLIP) per prompt. We set the same maximum generation budget of 6 images per prompt (1 initial and up to 5 additional rounds). For consistency, for correction amplification, we use α=1.09\alpha=1.09 for GenEval++, α=1.03\alpha=1.03 for GenEval2, and α=1.13\alpha= 1.13 GenEval.

GenEval Results

SettingOverall Accu.(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT82.01.140
Verifier guided Resampling85.01.511.51
Best of N (N = 6)85.066
Best of N (N = 2)84.022
Ablation
RECAF(always Fallback + Amplification)89.01.291.29
RECAF89.01.331.31

GenEval++ Results

SettingOverall Accu.(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT55.41.290
Verifier guided Resampling61.12.392.39
Best of N (N = 6)63.666
Best of N (N = 3)58.633
Ablation
RECAF(always Fallback + Amplification)60.02.352.34
RECAF61.12.422.41

GenEval2 Results

SettingSoft TIFA(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT36.62.250
Verifier guided Resampling37.84.754.75
Best of N (N = 6)37.866
Best of N (N = 5)34.755
Ablation
RECAF(always Fallback + Amplification)38.84.554.5
RECAF40.14.684.64

Comparison to verifier-guided resampling and Best of N. RECAF matches or outperforms verifier-guided resampling across all three benchmarks with comparable or fewer generation and verification calls. Under comparable generation budgets, RECAF also consistently outperforms Best-of-N. Increasing the budget to N=6, Best-of-N outperforms RECAF only on GenEval++. However, Best-of-N (N=6) requires six image generations and six verifier calls for each prompt, whereas RECAF achieves competitive or superior performance with substantially fewer calls.These results demonstrate that RECAF offers better performance under matched generation budgets.

Compared to Always Fallback + Amplification ablation, RECAF matches its performance on GenEval and outperforms it on the more challenging benchmarks, GenEval++ and GenEval2, with comparable inference costs. These results show that RECAF effectively integrates external verification with the UMM's intrinsic reflection, rather than replacing it entirely.

AuthorsOpenReview ↗

Reply to Reviewer Gbub - Part 3

We further evaluate Uni-CoT with a larger maximum refinement budget (20 rounds instead of 5) and compare it against RECAF.

BenchmarkSettingScore (↑)Images/Prompt (↓)Latency (s) (↓)
GenEvalUni-CoT (5 rounds)82.01.1464.5
Uni-CoT (20 rounds)82.01.2069.0
RECAF (5 rounds)89.01.33105.7
GenEval++Uni-CoT (5 rounds)55.41.2976.4
Uni-CoT (20 rounds)58.91.3177.5
RECAF (5 rounds)61.12.42224.6
GenEval2Uni-CoT (5 rounds)36.62.25156.5
Uni-CoT (20 rounds)36.82.69193.6
RECAF (5 rounds)40.14.68481.6

The results above show that allowing more refinement rounds does not necessarily lead Uni-CoT to perform better than RECAF. With a 20-round budget, the model only slightly increases the number of generated images per prompt. The reason is that the model terminates early with false results, leaving much of the available round budget unused. This behavior is consistent with the reflection failure identified in our paper: when the model incorrectly judges an image as satisfactory, simply increasing the maximum number of rounds cannot trigger further correction.

We also observe cases where Uni-CoT continues refining for multiple rounds, reaching the 20-round limit, yet still fails to correct the image. These cases illustrate the correction failure: additional editing attempts alone do not guarantee that the intended changes will be realized.

Overall, these results suggest that simply allowing more refinement rounds is insufficient to address the two identified failure modes we observed. RECAF instead intervenes at both stages by checking for potentially premature termination and strengthening the execution of correction instructions.

We have included the new baselines, along with the experimental configuration and inference-cost statistics, in Sec. 4.3, App. E, and App. G of the revised version.

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

The paper studies compositional image generation with self-reflective unified multimodal models, distinguishing missed errors during reflection from unsuccessful correction. It proposes RECAF, a training-free extension of Uni-CoT that checks proposed termination using SAM3 and BLIP, and amplifies correction-token keys and values during regeneration. It reports improvements on GenEval, GenEval++, and GenEval2, supported by component and K/V ablations. The evaluation demonstrates promising gains for the augmented system on one backbone, while leaving questions about verification feedback, inference budgets, and the contribution of intrinsic reflection.

Critical strengths
  • Separating premature acceptance from unsuccessful correction provides a useful framework for understanding failures in iterative image refinement and identifying intervention points.
  • The reported gains across three benchmarks and the improvement from amplification alone suggest practical value beyond external verification. The component ablations provide useful, although incomplete, evidence about where the gains originate.
Critical weaknesses
  • The main comparisons do not adequately control access to external verification feedback. The fallback checks object, count, color, and spatial constraints closely matching GenEval’s scoring targets, potentially using benchmark metadata, and triggers further editing when violations remain (Sec. 3.2; App. B.2). This gives RECAF additional feedback and correction opportunities. Existing ablations help, but do not establish superiority over verifier-guided resampling or selection under comparable budgets. Moreover, fallback-only matches reflection-with-fallback, while always-fallback plus amplification is untested. The headline gains therefore do not establish a clear advantage from the proposed reflection strategy.
  • The headline results remain difficult to reproduce and assess reliably. The amplified layers and important verification-to-instruction rules are unspecified. Appendix B.1 leaves GenEval’s amplification scale as “best performing,” while Table 6’s listed sweep reaches only 88–89%, versus the reported 90%. The absence of a stated independent tuning protocol and run uncertainty further limits confidence in the gains.
Questions for authors
  1. Does inference use any official benchmark evaluator components or outputs? Which metadata fields construct the fallback requirements, and can all of them be obtained from the prompt alone?
  2. Under the same verifier access and comparable generation costs, how does RECAF compare with verifier-guided resampling/selection and always-fallback plus amplification? Please report actual generation and verifier-call counts.
  3. Which exact configuration produces GenEval’s 90%, why is it absent from the reported scale sweep, and how were hyperparameters selected independently of evaluation?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer WRRa - Part 1

We thank the reviewer for the time invested in the review and suggestions. We appreciate the reviewer’s recognition that our work “provides a useful framework for understanding failures” and their positive evaluation of the “promising gains” demonstrated by our experimental results. We address the reviewer’s concerns in detail below.

W1&Q1:

To clarify, we do not use official benchmark evaluator components or outputs. We used the metadata to construct the requirement sets for the verifier. However, we found that such metadata is not necessary. The same requirement sets can be extracted directly from prompts using a rule-based parser, as in [1].

To demonstrate this, we employ a rule-based parser [1]. To evaluate the accuracy of the parser-extracted requirement set against a benchmark metadata-derived requirement set, we follow [2] and calculate the requirement-set exact-match accuracy. The results from all three datasets show that the parser-derived requirement set achieves 100% exact-match accuracy against the metadata requirements, removing the need for the metadata from benchmarks. We revised our paper accordingly in App. B.2.

[1] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating Semantically Precise Scene Graphs from Textual Descriptions for Improved Image Retrieval. In Proceedings of the Fourth Workshop on Vision and Language, Portugal. Association for Computational Linguistics. 2015

[2] Anderson, Peter, et al. Spice: Semantic propositional image caption evaluation. European conference on computer vision. 2016

AuthorsOpenReview ↗

Reply to Reviewer WRRa - Part 2

W1&Q2:

Compare with verifier-guided resampling/selection: To disentangle the benefits of additional generation attempts, we perform comparisons against verifier-guided resampling and Best-of-N selection.

Verifier-guided resampling: Each generated image is evaluated using the feedback from the Fallback verifier. If violations are detected, the model generates a new image from scratch using a different random seed. This process repeats until an image satisfies all verifiable constraints or the maximum generation budget is reached. If no candidate passes, we select the one with the fewest detected violations.

Best-of-N selection: The model independently generates N candidate images using different random seeds. The candidate images are evaluated using the feedback from the Fallback verifier, and the image with the highest verifier-derived constraint satisfaction is selected as the final output. We choose N to match the maximum number of generated images (1 initial+ 5 rounds) we set in RECAF, or set the budget to match RECAF (average 2 for GenEval, 3 for GenEval++, 5 for GenEval2).

As an additional ablation, we evaluate Always Fallback + Amplification, where the model always applies the fallback and amplification. We refer to it as RECAF(always Fallback + Amplification) below.

We compare across three datasets and report overall accuracy, the average number of generated images per prompt, and the average number of verifier calls (SAM+BLIP) per prompt. We set the same maximum generation budget of 6 images per prompt (1 initial and up to 5 additional rounds). For consistency, for correction amplification, we use α=1.09\alpha=1.09 for GenEval++, α=1.03\alpha=1.03 for GenEval2, and α=1.13\alpha= 1.13 for GenEval.

GenEval Results

SettingOverall Accu.(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT82.01.140
Verifier guided Resampling85.01.511.51
Best of N (N = 6)85.066
Best of N (N = 2)84.022
Ablation
RECAF(always Fallback + Amplification)89.01.291.29
RECAF89.01.331.31

GenEval++ Results

SettingOverall Accu.(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT55.41.290
Verifier guided Resampling61.12.392.39
Best of N (N = 6)63.666
Best of N (N = 3)58.633
Ablation
RECAF(always Fallback + Amplification)60.02.352.34
RECAF61.12.422.41

GenEval2 Results

SettingSoft TIFA(↑)Images / Prompts(↓)Verifier Calls/Prompts(↓)
Baseline
UniCoT36.62.250
Verifier guided Resampling37.84.754.75
Best of N (N = 6)37.866
Best of N (N = 5)34.755
Ablation
RECAF(always Fallback + Amplification)38.84.554.5
RECAF40.14.684.64

Comparison to verifier-guided resampling and Best of N. RECAF matches or outperforms verifier-guided resampling across all three benchmarks with comparable or fewer generation and verification calls. Under comparable generation budgets, RECAF also consistently outperforms Best-of-N. Increasing the budget to N=6, Best-of-N outperforms RECAF only on GenEval++. However, Best-of-N (N=6) requires six image generations and six verifier calls for each prompt, whereas RECAF achieves competitive or superior performance with substantially fewer calls. These results demonstrate that RECAF offers better performance under matched generation budgets.

Compared to Always Fallback + Amplification ablation, RECAF matches its performance on GenEval and outperforms it on the more challenging benchmarks, GenEval++ and GenEval2, with comparable inference costs. These results show that RECAF effectively integrates external verification with the UMM's intrinsic reflection, rather than replacing it entirely. We have incorporated the new baselines, along with the experimental configuration and inference-cost statistics, in Sec. 4.3 and App. E of the revised version.

W2&Q3:

The reported 90% on GenEval uses best performing category-specific amplification scales: α=[1.03,1.03,1.03,1.15,1.15,1.05]\alpha=[1.03,1.03,1.03,1.15,1.15,1.05] for the GenEval categories, respectively. Notably, a single global scale can already achieve 88-89%, as shown in Tab.15. We have clarified this configuration in Appendix B.1. The scales were selected using a subset of benchmark prompts and then fixed for full-benchmark evaluation. To address evaluation uncertainty, we additionally evaluated RECAF on GenEval with three random seeds with a single global scale (α=1.13\alpha=1.13). Results are below. GenEval++ results in App. I

CategorySingle Obj.Two Obj.CountingColorsPositionColor Attri.Overall
RECAF99.3±0.695±2.680.3±5.593±2.687.3±2.978±189±0.6
AuthorsOpenReview ↗

Reply to Reviewer WRRa - Part 3

W2 - Implementation details on verification-to-instruction: We thank the reviewer for pointing out the missing implementation details. We apply the correction amplification in all layers. While Sec. 3.2 and Fig. 5 describe how verifier-detected violations trigger correction, we acknowledge that the instruction-construction procedure was not sufficiently specified. During Fallback, the verifier produces structured constraint violations, which are converted into textual correction instructions through a template-based procedure, without introducing an additional model. Specifically, we use the following templates (simplified for brevity):

  • Count/Presence: "Add ⟨N⟩ ⟨attribute⟩ ⟨object⟩" for missing objects, or "Remove ⟨Δ⟩ ⟨object⟩" when the detected object count exceeds the required count.
  • Spatial relation: "Move ⟨object⟩ into ⟨target region⟩ so that it is ⟨relation⟩ ⟨reference object⟩."
  • Attribute: "Change the color of ⟨object⟩ to ⟨required color⟩, keeping its shape, texture, size, and position unchanged."

Here, ⟨·⟩ denotes slots filled from the prompt requirements and verifier outputs, while Δ denotes the number of excess instances. Spatial locations are derived from the verified image layout. Each instruction additionally includes preservation constraints to keep unrelated objects, the background, and the overall visual style unchanged.

For example, if the prompt requires one yellow TV remote but the verifier detects none, the resulting instruction is: "Add one yellow TV remote in an empty region of the image, without covering existing objects. Keep other required objects and the overall visual style unchanged."

We have added the implementation details in App.B.3 of the revised version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We sincerely thank you for your constructive feedback.

Our submission is driven by the underexplored research question: Why do self-reflective UMMs struggle with compositional image generation, and how can these failures be mitigated? Unlike existing approaches that primarily rely on additional training data, we systematically investigate the underlying causes of these failures and propose RECAF to address them directly.

We first summarize the strengths recognized by Reviewers. We appreciate Reviewer WRRa’s recognition of our useful framework for understanding failure and promising gains; Reviewer Gbub’s acknowledgment of RECAF’s potential applicability to existing self-reflective UMMs and stable gains without additional training; and Reviewer x8WA’s recognition of RECAF as an effective training-free approach that offers a useful perspective on the limitations of unified multimodal models.

We posted our detailed response and new experiments to address all concerns on 13. November. Following our response, Reviewer x8WA confirmed that their concerns have been addressed and raised their rating from 2 to 3 (Weak Accept). Now we summarize how these revisions and experiments address the main concerns raised by the reviews:

Budget-Matched Comparisons (WRRa, Gbub, x8WA) . We added Best-of-N and verifier-guided resampling baselines, showing that RECAF consistently outperforms both under matched generation budgets across three datasets. RECAF also outperforms Uni-CoT even when Uni-CoT is allocated a larger generation budget.

Benchmark Metadata (WRRa, x8WA). We clarified that RECAF does not rely on benchmark-specific detectors or metadata. A simple rule-based parser recovers the same requirement sets as benchmark metadata, confirming that such metadata is unnecessary.

Correction Amplification Analysis (Gbub, x8WA). Further analysis shows that RECAF improves the instruction-following rate on GenEval from 32% to 51% using correction amplification. Following reviewer x8WA’s suggestion, we also implemented three additional correction-enhancement baselines, all of which RECAF outperforms in overall accuracy, further validating its effectiveness.

Methodological novelty (x8WA). Our key contribution is a failure-aware correction framework that identifies and addresses two distinct bottlenecks in iterative generation. To demonstrate its broader applicability beyond compositional image generation, we evaluate RECAF on two reasoning benchmarks, WISE and T2I-ReasonBench. Averaged over three random seeds, RECAF improves scores from 71.2 to 73.1 on WISE and 52.5 to 54.9 on T2I-ReasonBench, demonstrating its broader potential.

We have revised the submission and marked all additional content in blue. During the rebuttal period, we carefully addressed the reviewers’ concerns through clarifications and additional experiments. Although we unfortunately did not have the opportunity for further discussion with the other two reviewers, reviewer x8WA, who raised similar concerns, increased their rating to Weak Accept following our rebuttal. Overall, these revisions and clarifications further demonstrate the significance and broader potential of our contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.