acceptodds
Withdrawn from ICLR 2027

R3: Rubric-Guided Region Reward for Precise Text-to-Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern text-to-image (T2I) models support complex, structured conditions involving multiple entities, attributes, text, and spatial layouts, yet existing image rewards remain insufficiently granular. Global rewards yield a single score for the entire image, and recent rubric-based rewards improve semantic granularity but judge each criterion against the whole image. We introduce **R3** (**R**ubric-guided **R**egion **R**eward), which reformulates image evaluation as an explicit target–evidence matching problem. R3 first structures the generation intent into entity-bound, coarse-to-fine rubrics and unambiguous grounding labels; then independently records what is present in the image through visual observations; finally, it matches visual evidence with the target specification while explicitly accounting for the spatial consistency of the observed targets. This design turns the reward estimation from direct holistic judgment into grounded verification: structured intent in, grounded credit out. We also introduce **GranuBench**, featuring high-quality contemporary generations, complex structured specifications, and explicit spatial targets to systematically evaluate reward granularity. On GranuBench, R3-based Gemini reaches a preference accuracy of 90.8%, outperforming direct-Gemini (80.7%) and existing reward models by a substantial margin. We further construct **R3Traj-100K**, containing 100K high-quality multi-stage R3 supervision trajectories, each recording structured target specifications, localized visual observations, and target–evidence assessments. Using R3Traj-100K, we distill the Gemini-instantiated R3 pipeline into an open 27B VLM with stage-specialized adapters, achieving state-of-the-art performance on multiple reward benchmarks. R3 can also improve image quality at test time by enabling precise region-targeted refinement. Together, these results demonstrate that finer reward granularity benefits not only evaluation but also inference- and training-time optimization.

Primary area:
generative models
Keywords:
Image Generation, Reward Signal

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YwdqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces R3, a framework that decomposes text-to-image evaluation into target structuring, independent visual observation, and target–evidence matching. Entity-bound rubrics, frozen observations, and a separate spatial-matching channel produce region-level and image-level rewards. The authors also introduce GranuBench and R3Traj-100K and distill the pipeline into Open R3 using stage-specific LoRA adapters.

Critical strengths

[1] The framework is well motivated, with explicit information boundaries and deterministic reward aggregation. Entity-level feedback is useful for traceable evaluation and targeted correction.

[2] Component ablations, external benchmarks, and a blinded human study provide meaningful evidence of practical usefulness.

Critical weaknesses

[1] The writing is occasionally verbose and convoluted, while key experimental details are insufficiently clear, making the evaluation protocol difficult to follow and reproduce.

Questions for authors

[1] The authors are encouraged to evaluate R3 for reward-guided T2I post-training and compare it with RubricRL and SpatialReward on DPG-Bench and GenEval, preferably using shared backbones such as SD3.5, FLUX.1-dev, or Qwen-Image under comparable experimental settings, with matched post-training algorithms, training data, and compute budgets.

[2] A sharper comparison with SpatialReward and RubricRM would help isolate the contribution of the information-separation design.

[3] Although R3 omits SAGE when no meaningful spatial target is available and uses a soft spatial penalty, how are multiple valid layouts handled once a target box is assigned? Please provide GranuBench statistics on target-box coverage and a human audit assessing whether spatial penalties are applied to valid alternative compositions.

[4] Equations (2) and (3), together with their accompanying explanations, are repetitive and could be consolidated for conciseness.

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

This paper introduces R3, a text-to-image reward framework that separates target structuring, visual observation, and target–evidence matching, combining entity-level semantic and visual-quality assessments with spatial consistency scores. The authors introduce GranuBench with 1,000 human-labeled preference pairs and 100K training trajectories, and distill the pipeline into the 27B Open R3 model. R3 achieves strong preference accuracy on GranuBench and three external benchmarks, while single-round refinement experiments and blinded human evaluations demonstrate the practical value of region-grounded feedback.

Critical strengths
  1. Clear and traceable design. Separating target specifications from visual observations and binding judgments to entities enables precise error diagnosis.
  2. Practical gains supported by human evaluation. R3 improves clean target-rubric repair from 41.0% to 60.2%, and edited images are preferred over their originals in 94.08% of blinded comparisons.
Critical weaknesses
  1. Spatial target provenance is unclear. SAGE raises Open R3 accuracy from 63.8% to 90.0%, yet the construction and freezing of target boxes are insufficiently documented. Since annotators inspect candidate images before defining the specifications, independence from candidate content and preference labels needs clarification.
  2. Model-generated boxes may penalize valid layouts. For prompts without explicit spatial constraints, a single predicted layout may exclude equally valid compositions. Separate results for spatially constrained and unconstrained prompts, together with box-sensitivity analysis, are needed.
Questions for authors

Q1. How are target boxes constructed and frozen across GranuBench, the development and diagnostic sets, and external benchmarks? Do they use candidate images or preference labels? Please clarify “preference-anchored,” “reference-anchored,” and train–test deduplication. Q2. Can you report results with targets constructed solely from the original prompts, before viewing candidate images? Please separate prompts with and without explicit layout constraints and assess sensitivity to alternative valid layouts. Q3. Table A6 reports 9/100 false positives, while the text states 1/11. Which is correct? Please clarify sample sizes and provide human-validated precision, recall, and confidence intervals. Q4. Can you isolate the contributions of entity binding, localization, and frozen observations using the same model and comparable inference budgets? Please provide key prompts, training settings, and inference costs. Q5. Are the small reductions in non-target damage statistically significant? Please clarify target selection and provide prompt-clustered confidence intervals, ideally with direct human comparisons between R3 and rubric-guided outputs.

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

Motivation: The paper addresses the semantic–visual granularity mismatch in T2I evaluation.

Proposed method: R3 is a reward formulation/pipeline. It first structures the target by decomposing the prompt into entity-level rubrics and target bounding boxes. A VLM then describes each entity without access to the rubrics. Finally, the target and the evidence are matched purely in text space. The final score combines semantic satisfaction, local visual quality, and spatial consistency (SAGE).

The contributions also include GranuBench, which consists of 1,000 preference pairs built from real-world requests collected from online communities. Each request is first expanded into a detailed caption by Gemini-3.7-Flash, and candidate images are generated with SOTA T2I models. Three annotators label the entities, grounding queries, and rubrics, which are then converted into structural prompts. Three imaging professionals make the preference judgments, and the final label is decided by majority vote. The benchmark includes explicit target bounding boxes. R3Traj-100K and Open R3: Complete trajectories are generated with the Gemini-based R3 and distilled into Qwen3.8-27B.

Conclusion: The paper concludes that finer reward granularity benefits both evaluation and generation.

Critical strengths

The problem is clearly defined, and both the motivation and the solution make sense. The structured, entity-level outputs can be used directly for targeted editing, which is supported by the experiments and the blinded human evaluation.

Critical weaknesses
  1. Methodology Novelty: Most individual components have close antecedents, and the novelty appears to lie primarily in their integration and information-flow design rather than in a new learning algorithm. The information isolation substantially reduces target-conditioned false affirmation, but the paper does not show that it clearly improves preference ranking. The headline improvement critically depends on SAGE. The 90.8% accuracy relies heavily on spatial information: without SAGE, accuracy drops to about 64%. The experiments therefore only show that R3 can recover human rankings well when the evaluator is given target boxes that may be correlated with the reference preference.
  2. Experiment Solidness: The self-built benchmark is closely aligned with the method. GranuBench explicitly provides entities, rubrics, and bounding boxes, which are exactly the inputs R3 requires. Conventional reward models cannot fully exploit this information, so the comparison is not fair.
Questions for authors
  1. Inference cost: images with many entities require multiple rounds of observation, rubric construction, and judgment. Could the authors report compute metrics such as latency and the number of API calls per image?
  2. The abstract claims that the approach benefits training-time optimization. However in the paper, R3 is used only as an evaluator and for test-time feedback; the paper never uses it to directly train any T2I model (as opposed to the evaluator model itself). A good evaluator is not necessarily a good RL reward: the paper only checks agreement with human preferences and does not examine reward hacking. Where did the training-time optimization claim come from?

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

Related papers

Discussion (0)

Sign in to comment.