acceptodds
Under review as a conference paper at ICLR 2027

SCOPE: Semantic-Bounded Credit Optimization for Prompt Engineering in Text-to-Image Generation

Abstract

Training a language model to rewrite prompts can improve text-to-image (T2I) generation while keeping the image generator frozen. In common reinforcement-learning formulations, a single image-level reward produces one trajectory-level advantage that is applied across the tokens of a rewritten prompt. This creates a credit-assignment gap: the units that users compose and that T2I systems must bind—objects, counts, attributes, and relations—are not the units that receive policy-gradient credit. We present SCOPE, a GRPO-style objective that uses semantic spans to structure this token-level policy update. For each rewritten prompt, SCOPE aligns generated spans with tokenizer positions, averages current-versus-old policy log-probability shifts within each span, and converts these shifts into detached weights normalized by valid-token count. The resulting weights modulate the token-level clipped policy surrogate, while the reward remains at the trajectory level. SCOPE therefore does not estimate the causal contribution of individual spans to the image reward; it provides a span-aware reweighting of policy updates. Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE improves over vanilla GRPO on GenEval and T2I-CompBench, with mean relative gains of 9.48% and 7.62%, respectively. These results show that span-aware policy-update reweighting improves compositional alignment in the evaluated frozen-renderer settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text-to-image generation, prompt optimization, reinforcement learning, credit assignment, GRPO

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_XSg3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCOPE, a reinforcement learning framework for prompt rewriting in T2I generation.

Standard prompt rewriting RL gives a single image-level reward equally across all generated tokens, which treats split subwords as independent update targets. SCOPE asks the rewriter language model to produce semantic phrases separated by dividers during generation. It averages the policy log-probability changes within each phrase and uses them as detached weights to adjust the token-level policy loss. The authors test SCOPE with Qwen3-8B across Qwen-Image, SD3, and FLUX.1-dev on GenEval and T2I-CompBench.

Critical strengths
  1. The span-level reweighting method introduces no extra visual reward models or extra rendering cost. The rewriter language model generates phrase separators together in a single pass, and the span weights reuse already computed log-probabilities without needing an external text parser.

  2. The training pipeline includes a reliable fallback mechanism that protects training stability. When the model generates wrong separators, misaligned tokens, or cut-off outputs, the loss automatically falls back to standard uniform-token GRPO, with recorded fallback rates staying below 0.6% across all runs.

  3. The paper is well-written and easy to follow.

Critical weaknesses
  1. The performance gains over standard GRPO are marginal and show clear drops in some categories. On GenEval, the overall improvement on Qwen-Image is only 0.04, while the color score drops from 0.94 to 0.93. Across all three backbones on T2I-CompBench, the average score gains are only around 0.04, which is very small considering the randomness in diffusion generation.

  2. The proposed span weighting only reflects the size of policy changes without direct credit assignment to the image reward. A phrase with a larger change between current and old policies does not necessarily contribute more to visual alignment. The method scales token weights based on probability changes without checking true visual cause, leaving the credit assignment problem unsolved.

  3. The span weights are limited to a very narrow range between 0.8 and 1.2. This tight bound makes the weight change act like a mild smoothing factor on token gradients. The paper lacks tests on different ranges, so it is unclear whether the performance gain comes from real phrase weighting or simple gradient smoothing.

  4. The evaluation relies entirely on automated detector scores without human verification. The paper lacks blind human studies to verify if these small score changes lead to noticeably better images.

Questions for authors

Please kindly refer to the weaknesses.

6 comments

Thank you for your careful review and constructive suggestions. We respond to each point below.

Q1: Are the gains stable under diffusion sampling?

A1: The gains are stable under diffusion sampling because we control sampling randomness during both training and evaluation. During training, all members of each rollout group use the same image-generation seed, reducing interference from different sampling noise within a group. Different rollout groups and training steps use different seeds, so training does not rely on a single fixed seed. During evaluation, each GenEval prompt is used to generate images with 4 different seeds, and each T2I-CompBench prompt with 10 different seeds. Benchmark scores are computed from these generated images. These training and evaluation protocols are described in Appendix A.13 of the submitted paper.

Q1.a: Why does the Colors score decrease from 0.94 with GRPO to 0.93 with SCOPE?

A1.a: SCOPE uses policy-change signals to adjust span reweighting, so its gains depend on how much different semantic combinations help the generation task. Our inspection of rewritten prompts shows that the effective combinations learned by the model mainly concern counting layouts and the coordination of absolute positions with relative relations. These descriptions directly reinforce counting and spatial constraints, whereas color requirements depend more on correctly associating colors with objects and may benefit less from the same kinds of layout descriptions. In addition, Qwen-Image already has strong color-generation ability: after GRPO optimization, its Colors score is 0.94, leaving less room for further improvement.

Q2: Why use policy changes to construct span weights, and which aspect of credit assignment does this improve?

A2: Policy changes use an update signal already available in GRPO training. SCOPE averages this signal within each semantic phrase, so tokens expressing the same requirement share an update coefficient. This refines the granularity of policy updates, aligning them with the semantic structure of the prompt without requiring an additional VLM or phrase-level reward model.

We agree that probability changes between the current and old policies reflect policy-update signals and do not directly measure a semantic span’s contribution to the image reward or visual alignment. SCOPE does not solve causal credit assignment at the span level. Instead, it addresses a narrower issue: an image-level advantage would otherwise be applied uniformly to every token in a rewrite. SCOPE pools policy changes within semantic spans and assigns a shared coefficient to the tokens in each span, aligning the granularity of updates with the semantic structure of the prompt.

Q3: Do semantic boundaries matter beyond general gradient smoothing, and how sensitive are the results to the clipping range?

A3: The partition ablation shows that semantic grouping matters, while the range ablation shows that moderate clipping performs best. These experiments examine different factors.

To test the partitioning scheme, Figure 3(b) compares semantic spans with token-local, random-span, and fixed-window partitions on Qwen-Image.

Partitioning schemeGenEval Overall
Token-local0.89
Random spans0.86
Fixed 3-token windows0.87
GRPO0.91
SCOPE semantic spans0.95

Semantic spans score 0.95, while the other partitioning schemes score between 0.86 and 0.89. Thus, grouping arbitrary adjacent tokens does not reproduce the result achieved by semantic spans.

To test the strength of reweighting, we hold the other settings fixed on FLUX.1-dev and vary the raw clipping range before normalization.

ExperimentRaw span-weight clipping rangeGenEval Overall ↑SingleTwoCountingColorsPositionAttribute
E1: Reweighting disabled[1.0, 1.0]0.770.980.860.740.830.620.56
E2: Weaker reweighting[0.9, 1.1]0.810.990.860.770.860.800.57
E3: Default setting[0.8, 1.2]0.841.000.890.830.890.840.60
E4: Stronger reweighting[0.6, 1.4]0.800.990.870.750.850.790.57

The default range achieves an Overall score of 0.84, compared with 0.77 when reweighting is disabled, 0.81 with the narrower range, and 0.80 with the wider range. Thus, both the choice of update unit and the strength of reweighting affect performance.

Q4: Are the gains perceptible to people, and was a blind evaluation conducted?

A4: Human raters can perceive the improvement: SCOPE achieves preference rates of 74.28% over GRPO and 88.26% over the original baseline. The completed study is a paired comparison on FLUX.1-dev.

We conducted a blind evaluation: 5 raters independently assessed all 553 GenEval prompts, with 4 seeds per prompt, yielding 2,212 matched image pairs for each comparison.

Raters used the original prompt as the reference and made a forced-choice judgment with no ties. If only one image met the original prompt’s requirements, they selected that image. If both images met the requirements, they compared visual quality. If neither image met the requirements, they considered both instruction following and visual quality.

GenEval dimensionImage pairsSCOPE preference over GRPOSCOPE preference over the original baseline
Single object32060.25%87.75%
Two objects39668.33%90.51%
Counting32082.00%91.56%
Colors37676.91%81.54%
Position40087.00%96.60%
Color attribution40070.00%81.80%
Overall2,21274.28%88.26%

Preference rates are averaged across raters, and Overall is weighted by the number of image pairs. SCOPE’s preference rates over GRPO reach 82.00% for Counting and 87.00% for Position. Thus, the gains on compositional generation are also reflected in human judgments.

Thanks for the response. A3 answers my question on the clipping range. The human study in A4 also mostly addresses my concern on human evaluation.

Two main concerns are still not resolved:

  1. A2 agrees that SCOPE does not solve causal credit assignment, and says tokens in each semantic span share an update coefficient. But in uniform-token GRPO, all tokens already have weight 1.0, so each span is already consistent inside. SCOPE gives different weights to different spans, yet there is no clear reason why a span with a larger policy log-probability shift should get a larger weight. Calling it 'Credit Optimization' and 'SCOPE credit assignment' is an overclaim.

  2. A1 only repeats Appendix A.13 and gives no standard deviations or multi-run variance. So it is unclear if the ~0.04 gains on T2I-CompBench are much larger than renderer noise.

So my main concerns on the method motivation and score stability remain.

Reviewer_3rdrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work trains Qwen3-8B with GRPO to rewrite prompts for frozen text-to-image backbones (Qwen-Image, SD3, FLUX.1-dev). The authors argue that standard GRPO broadcasts a single scalar advantage Ag,iA_{g,i}, obtained from the rendered image, uniformly to every subword token of the rewrite, whereas what actually matters in text-to-image generation are semantic phrases such as "four red handbags" or "on white pedestals."

SCOPE does three things:

  1. The rewriter emits the rewrite directly as |-delimited semantic phrases (<answer>span 1 | span 2 | ...</answer>); the | characters are removed before the prompt is passed to the image model.
  2. Each group samples N−1N-1 rewrites plus the original prompt as a reference, and all group members share the same rendering seed. When the original prompt's score is below a threshold τ\tau, the advantage is rg,i−bgr_{g,i}-b_g; otherwise standard group normalization is used (Equation 3).
  3. For each span, the per-token log-probability shift between the current and old policy is averaged into bg,i,sb_{g,i,s}, passed through clip(exp⁡(βbg,i,s),0.8,1.2)\text{clip}(\exp(\beta b_{g,i,s}),0.8,1.2) and normalized by token count (with stopgrad), and then multiplied onto the advantage for that span.
Critical strengths
  • The boundaries of the method are stated honestly. Both the abstract and the conclusion state clearly that the method does not estimate each span's causal contribution to the image reward; no span-level critic or counterfactual re-rendering is needed, and rendering cost is barely increased.
  • Comparisons are run on three mainstream open-source backbones and on both GenEval and T2I-CompBench, with the gains concentrated on known weak categories such as counting and spatial position.
  • Fixing the rendering seed within a GRPO group and including the original, unedited prompt in the group as a reference is a practical design for stabilizing RL training on text-to-image generation.
Critical weaknesses
  • Location: Table 1 (lines 335, 344), Appendix A.15 (lines 1001–1007) In Table 1, the reproduced RePrompt† row for Qwen-Image (0.98 / 0.87 / 0.77 / 0.85 / 0.62 / 0.49 / 0.76) is identical, number for number, to the RePrompt row for FLUX.1-dev taken from the original paper. One is a reproduction on a different backbone, the other a citation, yet all seven numbers match; this is unlikely to be a coincidence, especially since Appendix A.15 goes out of its way to state that the Qwen-Image results were not transferred from other backbones. I checked Table 1 of the original RePrompt paper (Wu et al., 2025c): these seven numbers are exactly its reported FLUX.1-dev result (the "+Ours (train w/ FLUX)" row), which suggests that the Qwen-Image RePrompt† row reuses the published FLUX.1-dev result rather than a new reproduction. Moreover, this reproduced result (0.76) is clearly below the unmodified Qwen-Image Baseline (0.87), whereas RePrompt is above Baseline on both SD3 and FLUX.1-dev. Please check whether this is a copy-paste error, and provide the configuration and per-category raw results of the Qwen-Image RePrompt reproduction. Until this is clarified, I cannot treat that row as a valid comparison.

  • Location: Section 3.3, Equations (6)–(8), Appendix A.2 Table 3 The span weights may have very little effect. Δg,i,t\Delta_{g,i,t} is simply the log of the importance ratio; on the first gradient update of each rollout batch, πθ=πold\pi_\theta=\pi_{\text{old}}, so all weights equal 1 and SCOPE is exactly identical to GRPO. The weights only deviate from 1 on later update steps, when a single batch of rollouts is reused over multiple updates. Even then, the weights are clipped to [0.8, 1.2] and normalized, so within one rewrite the weights of different spans differ by at most a factor of 1.5. The paper does not report the number of updates per batch or the value of β\beta. A reweighting that is bounded, detached from the gradient, and equal to 1 on the first update of each batch is hard to reconcile with gains as large as +0.10 overall and +0.26 on position for SD3 on GenEval. At minimum, the actual distribution of the weights during training should be reported.

  • Location: Appendix A.9 (lines 884–889), Figure 3, Table 1 SCOPE and GRPO differ by more than span reweighting. In Appendix A.9, the SCOPE system prompt contains an additional instruction, absent from the GRPO prompt, asking the model to split the rewrite into semantic phrases; this alone can change the structure and quality of the rewrite. In Figure 3(c) on FLUX.1-dev, SCOPE with the z-norm advantage (0.832) and with the hybrid advantage (0.841) differ by only 0.009, while both differ from the GRPO row in Table 1 (0.76) by about 0.07–0.08 — a gap in which span reweighting and prompt format are confounded, and which the current experiments cannot separate. In Figure 3(a)(b), on Qwen-Image (panel (b) is not labeled, but its SCOPE score of 0.95 matches Qwen-Image), token-local (0.89), random spans (0.86) and fixed 3-token windows (0.87) are all below GRPO (0.91), and the paper does not say which prompt each of these variants used. This is hard to reconcile with the previous point: if span weights bounded in [0.8, 1.2] have as little effect as the analysis above suggests, replacing semantic spans with random or fixed-window spans should not fall 0.04–0.05 below GRPO. Either the weights matter much more than their bounds imply, or these variants differ from GRPO in other ways. The format reward rfmtr_{\text{fmt}} likely differs as well: per Appendix A.5 it checks for extractable spans, which the GRPO prompt does not ask for. The control that is most needed is SCOPE's span prompt and format reward with the same advantage rule and β=0\beta=0.

  • Location: Appendix A.5, Figure 4, Appendix A.9 system prompts, Table 1 The training reward and the evaluation metric come from the same evaluators. rt2ir_{\text{t2i}} is computed directly by the GenEval and T2I-CompBench evaluation models, and the training prompts come from the Flow-GRPO GenEval training pool and the official T2I-CompBench training split, so part of the gain may come from fitting the evaluators' preferences. The rewrites shown in Figure 4 also add content the user did not request, such as "each on a white pedestal" and "clean neutral background," which conflicts with the system prompt instruction not to add scene elements the user did not request. This has less bearing on the relative comparison between SCOPE and GRPO, since both are trained on the same reward, but it does affect whether the claim of improved compositional alignment generalizes. In addition, GenEval is close to saturation on Qwen-Image (the unmodified Baseline already reaches 0.87 and SCOPE reaches 0.95), which limits its discriminative power. I suggest adding at least one evaluation that was not used in training, such as DPG-Bench, GenAI-Bench, T2I-Keypoints-Eval [1] (which covers more challenging prompt-following categories and is judged per keypoint by an independent VLM), or a small human study. I also suggest showing failure cases in which images receive high evaluator scores but do not match the user's intent or have poor visual quality, and analyzing such score–quality discrepancies.

  • Location: Table 1 caption (lines 324–326), RePrompt rows in Tables 1 and 2, Appendix A.15 (lines 1006–1007), Related Work (lines 158–161) (Minor) The comparison with existing rewriting methods is not like-for-like, and an important piece of related work is missing. The SD3 and FLUX.1-dev RePrompt rows are copied from the original paper, but RePrompt uses Qwen2.5-3B (SFT followed by RL), a reward combining ImageReward and GPT-4V, and is trained only on FLUX.1-dev; its SD3 row is obtained by transferring the FLUX.1-dev-trained rewriter. SCOPE instead uses Qwen3-8B, computes rewards by rendering with each target backbone (Table 6), and directly uses the GenEval / T2I-CompBench evaluators as the reward. The Table 1 caption states that SCOPE improves over the strongest competing method by +0.09 on SD3, where that competitor is RePrompt; this gap cannot be attributed to SCOPE itself. I suggest either re-running RePrompt with the same base model, reward, and training data, or clearly marking these rows as reference only. In addition, the paper does not discuss PromptEnhancer [1], which also trains a rewriter with GRPO for a frozen text-to-image backbone, uses a reward model (AlignEvaluator) that scores 24 fine-grained keypoints and averages them, and reports GenEval and T2I-CompBench results on Qwen-Image. PromptEnhancer refines the evaluation dimensions of the reward while still using a sequence-level scalar reward in RL, whereas SCOPE refines the granularity of credit assignment, so the two are orthogonal. I suggest discussing their relationship in the related work and, where possible, including a comparison.

  • Location: Appendix A.2 Table 3, Table 1, Figures 4 and 5 (Minor) Table 3 is missing the number of updates per batch, minibatch size, β\beta, τ\tau, cc, the reward weights, the PPO clipping range, and the number of training steps; lines 782–784 also state that local launcher defaults may differ from the paper settings. The z-image rewrite drops single-object accuracy on SD3 and FLUX.1-dev from 0.98 / 0.99 to 0.58 / 0.53, which is a very large drop and needs an explanation. The baseline labeled gpt-image 2 in Figures 4 and 5 appears only in the column headers and caption; neither the main text nor the appendix says what it is or which prompt it was given.

[1] Wang et al. PromptEnhancer: Taming Your Rewriter for Text-to-Image Generation via Fine-Grained Reward. CVPR 2026 (arXiv:2509.04545). T2I-Keypoints-Eval is the benchmark released with this work.

Questions for authors
  1. In Table 1, the seven numbers in the Qwen-Image RePrompt† row and the FLUX.1-dev RePrompt row are identical and also match the FLUX.1-dev result reported in the original RePrompt paper. Was the published result reused? Please provide the raw results of the Qwen-Image RePrompt reproduction.
  2. What are the results on the three backbones when using SCOPE's span prompt, format reward, and the same advantage rule but setting β=0\beta=0? Do the GRPO rows in Tables 1 and 2 use the hybrid advantage of Equation (3)?
  3. How many gradient updates are performed per rollout batch? What is the value and sign of β\beta? What is the actual distribution of the span weights ww during training?
  4. Does SCOPE still improve over GRPO on an evaluator that was not used as the training reward (e.g., T2I-Keypoints-Eval) or in a human study? Can you show failure cases that receive high evaluator scores but do not match the user's intent?
  5. Please complete the missing hyperparameters in Table 3 and clarify the role of gpt-image 2 in the comparison.
  6. The RePrompt rows on SD3 and FLUX.1-dev differ from SCOPE in base model, reward, and training data. Can RePrompt be re-run under the same setting? How does SCOPE relate to PromptEnhancer, and can a comparison be provided?
3 comments

Thank you for your careful review and constructive suggestions. We address your questions below.

Q1: Were the published FLUX.1-dev results mistakenly copied into the RePrompt row for Qwen-Image?

A1: Yes, that row in the original table contains a copy error. Our actual RePrompt reproduction on Qwen-Image achieves an Overall score of 0.83. Thank you for identifying this issue. The RePrompt paper does not report results on Qwen-Image. The corrected comparison is as follows:

Evaluation on Qwen-ImageOverall
Original prompt0.87
RePrompt† reproduction0.83

The per-category results and reproduction settings are provided in A1.a.

Q1.a: What are the reproduction settings and per-category results for RePrompt on Qwen-Image, and why does it score below the original prompts?

A1.a: The largest decrease is in Position, from 0.79 to 0.69, followed by Attribute, from 0.75 to 0.66. Inspecting the reproduced rewrites, we found that they added descriptions but expressed spatial layouts less precisely; this observation is consistent with the Position results. We follow RePrompt’s specified SFT, RL, and reward configurations and use the same reward service, rendering seeds, and evaluation pipeline as GRPO and SCOPE.

MethodSingleTwoCountingColorsPositionAttributeOverall
Qwen-Image original baseline0.990.950.870.890.790.750.87
RePrompt† reproduction0.990.950.830.850.690.660.83

Q2: How many updates are performed per rollout batch, and what is β?

A2: Each rollout batch receives 1 optimizer update, with positive β = 1 and a lagged old-policy snapshot. The old policy is not reset to the current policy whenever a new batch is generated. The two policies match at initialization, so the initial weights are 1. Subsequent policy changes produce different span coefficients without repeatedly optimizing the same batch.

For span s, the computation is as follows:

  • bₛ is the mean Δ log π over the valid tokens in the span.
  • vₛ = clip(exp(βbₛ), 0.8, 1.2) is the coefficient before normalization.
  • We normalize vₛ so that the mean coefficient over content tokens is 1, and stop gradients through these weights.

β controls span sensitivity and is distinct from the KL coefficient of 0.01. The interval [0.8, 1.2] bounds the raw coefficients before normalization.

To test whether this modulation is effective, we hold the other settings fixed on FLUX.1-dev.

ExperimentRaw clipping rangeGenEval OverallSingleTwoCountingColorsPositionAttribute
E1: Reweighting disabled[1.0, 1.0]0.770.980.860.740.830.620.56
E2: Weaker reweighting[0.9, 1.1]0.810.990.860.770.860.800.57
E3: Default setting[0.8, 1.2]0.841.000.890.830.890.840.60
E4: Stronger reweighting[0.6, 1.4]0.800.990.870.750.850.790.57

Q3: What happens when β is set to 0 while the other SCOPE settings remain unchanged?

A3: Setting β to 0 reduces Overall from 0.95 to 0.92 on Qwen-Image, from 0.84 to 0.75 on SD3, and from 0.84 to 0.77 on FLUX.1-dev. To isolate span reweighting, this control keeps SCOPE’s system prompt, format reward, and hybrid advantage unchanged.

BackboneSettingSingleTwoCountingColorsPositionAttributeOverall ↑
Qwen-Imageβ = 00.990.980.910.920.900.840.92
Qwen-Imageβ = 11.000.990.940.930.950.880.95
SD3β = 00.980.880.610.860.590.600.75
SD3β = 11.000.900.790.890.830.630.84
FLUX.1-devβ = 00.980.860.740.830.620.560.77
FLUX.1-devβ = 11.000.890.830.890.840.600.84

At β = 0, the span coefficients become uniform weights. Comparing this setting with β = 1 under the same SCOPE configuration therefore tests the additional benefit of reweighting. Overall improves on all three backbones, with particularly clear gains in Counting and Position.

Q3.a: Do the GRPO results in Tables 1 and 2 use the hybrid advantage in Equation (3)?

A3.a: The GRPO results in Tables 1 and 2 do not use the hybrid advantage in Equation (3).

Q3.b: Which prompts and format rewards do the token-local, random-span, and fixed 3-token-window variants use, and why do they perform below GRPO?

A3.b: All partitioning variants use SCOPE’s rewriting system prompt and the same format reward. They differ in which tokens share a weight. The format reward checks whether a complete prompt can be extracted from <answer>...</answer>; it does not check whether semantic boundaries are correct.

To examine the choice of update unit, Figure 3(b) compares the following partitioning schemes on Qwen-Image.

Partitioning schemeGenEval Overall
Token-local0.89
Random spans0.86
Fixed 3-token windows0.87
GRPO0.91
SCOPE semantic spans0.95

Our explanation for their lower performance than GRPO is that token-local weights do not pool policy changes across tokens within a span. Random boundaries and fixed windows do not follow the prompt’s semantic boundaries: they can split objects from attributes, divide count descriptions or spatial relations, and group unrelated content into the same span, disrupting the semantic composition of the prompt during advantage-weighted updates.

Q4: Can an evaluator independent of training verify these gains, including perceptible improvements when Qwen-Image scores are near saturation?

A4: Yes. An independent blind human evaluation on FLUX.1-dev shows SCOPE preference rates of 74.28% over GRPO and 88.26% over the original baseline. To examine perceptible gains, we use 553 prompts with 4 seeds each, yielding 2,212 matched image pairs per comparison, independently assessed by 5 raters.

Raters use the original prompt as the reference and make a blind forced-choice judgment with no ties. If only one image meets the original prompt’s requirements, they select that image; otherwise, they judge visual quality and, where necessary, the degree of instruction following.

GenEval dimensionImage pairsSCOPE preference over GRPOSCOPE preference over the original baseline
Single object32060.25%87.75%
Two objects39668.33%90.51%
Counting32082.00%91.56%
Colors37676.91%81.54%
Position40087.00%96.60%
Color attribution40070.00%81.80%
Overall2,21274.28%88.26%

Q4.a: Do the added backgrounds and white pedestals narrow or change the user’s original request?

A4.a: Yes. Even when the objects and count specified in “four handbags” are preserved, white pedestals and a fixed arrangement narrow the range of scenes allowed by the request. These are additional scene choices, rather than details entailed by the original request.

SCOPE evaluates generated images against the original prompt. This rewards adherence to the requested objects, counts, attributes, and relations, but does not separately penalize every choice of an unspecified background.

Q4.b: Can you show failure cases with high automatic scores and explain the gap between these scores and user intent or image quality?

A4.b: High automatic scores for images that do not match user intent or have poor visual quality arise from evaluator misjudgments of objects, attributes, or relations, constituting hallucinations at the reward-model level. SCOPE focuses on organizing policy updates using the semantic structure of prompts. Improving the evaluator’s recognition and judgment is a complementary research direction outside the focus of this paper. We will add a set of 4 failure-case images to the appendix of the revised PDF.

Q5.a: Can RePrompt be retrained with the same model, data, and generation settings as SCOPE?

A5.a: This experiment requires retraining and evaluation across multiple image-generation backbones and cannot be completed within the rebuttal period. If the paper is accepted, we will add this comparison under matched settings in the camera-ready version.

Q5.b: How does SCOPE differ from PromptEnhancer, and what are the direct comparison results?

A5.b: SCOPE achieves 0.95 on GenEval with Qwen-Image, compared with 0.88 for PromptEnhancer’s released rewriter. The two methods address different aspects of reinforcement learning for prompt rewriting.

MethodMain design focus
PromptEnhancer [1]AlignEvaluator scores 24 fine-grained keypoints and aggregates them into a sequence-level reward to train a reasoning rewriter initialized with SFT
SCOPEShared coefficients within semantic spans modulate the effect of image-level advantages on policy updates

Reward design and update granularity are complementary. To compare deployed rewriters, we evaluate the released PromptEnhancer model and general-purpose LLMs with the same Qwen-Image configuration, GenEval prompts, and seeds.

MethodQwen-Image GenEval Overall ↑
Original prompt0.87
Officially released PromptEnhancer rewriter0.88
GPT 6 Astra: training-free rewriting0.88
Claude Fable 5.2: training-free rewriting0.88
Gemini 4 Carbon: training-free rewriting0.89
GRPO0.91
SCOPE0.95

The 0.88 is our evaluation result for the released PromptEnhancer model. The LLM baselines use a shared rewriting instruction adapted from Appendix A.9. SCOPE scores higher than all these rewriting baselines under the same generation and evaluation pipeline.

Q6: What are the actual training settings, including updates, batch size, β, τ, c, rewards, clipping, training duration, and launcher defaults? What is the total optimizer step count?

A6: The experiments use 1 optimizer update per rollout batch, a minibatch size of 8, and β = 1. The configuration is listed below.

ConfigurationSetting
Rewriting policyQwen3-8B
Frozen generatorsQwen-Image, SD3, FLUX.1-dev
Training data10,000 GenEval prompts and 4,200 T2I-CompBench prompts
Rollout group8 members: 7 rewrites and 1 original prompt
Minibatch size8
Training processes1
Gradient accumulation1
Optimizer updates per rollout batch1
Total optimizer steps1,775
Training duration1 epoch
Maximum generation length512 tokens
Learning rate0.000001
Span sensitivity β1
Raw span coefficient clipping[0.8, 1.2]
Hybrid advantage threshold τ1
Advantage clipping c3
Numerical stability term0.000001
Image reward weight1
Format reward weight1
KL coefficient0.01
PPO probability-ratio clipping[0.8, 1.2]

The experiments use TRL 0.24.0. Span coefficient clipping and PPO ratio clipping both use [0.8, 1.2], applied to span coefficients and probability ratios, respectively.

Q6.a: Why does z-image rewriting reduce the single-object scores on SD3 and FLUX.1-dev?

A6.a: Detailed expansion introduces scene content beyond the requested objects, reducing adherence to single-object requirements. The z-image instructions expand composition, lighting, materials, and spatial depth, but do not maintain a balance between instruction following and aesthetics. The measured changes are as follows:

GeneratorOriginal single-object scoreAfter z-image rewriting
SD30.980.58
FLUX.1-dev0.990.53

Q6.b: What is the gpt-image 2 column in Figures 4 and 5, and which prompts does it receive?

A6.b: These images are generated by feeding the original GenEval prompts directly to gpt-image 2, without applying our prompt-optimization pipeline. The column serves as a direct-generation reference from a closed-source image model for qualitative comparison with images generated from GRPO- and SCOPE-optimized prompts.

Figure columnPrompt supplied to the generator
gpt-image 2Original GenEval prompt
GRPOPrompt rewritten by GRPO
SCOPEPrompt rewritten by SCOPE

The figures show visual results for counts, attributes, and spatial relations. The closed-source model column serves as a qualitative reference.

Reviewer_36DZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SCOPE is a method for training an LLM to rewrite prompts for text-to-image generation while keeping the image generator unchanged. It addresses a mismatch between image-level feedback and token-level training: an image receives one overall score, but important requirements, such as object counts, colors, and spatial relationships, are expressed through meaningful phrases. SCOPE asks the model to generate both the rewritten prompt and its phrase boundaries. It then calculates a shared training weight for the tokens within each phrase based on changes in their generation probabilities. These weights adjust how strongly different parts of the prompt are updated; they do not measure each phrase’s contribution to the image score. Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE outperforms GRPO, with average relative improvements of 9.48% on GenEval and 7.62% on T2I-CompBench.

Critical strengths
  1. Consistent improvements across different image generators. SCOPE improves GenEval and T2I-CompBench scores over GRPO on all three tested backbones. The gains are particularly clear for counting and spatial relationships, which directly match the paper’s goal of improving compositional alignment.
  2. A practical use of semantic structure. SCOPE incorporates phrase-level structure into the existing GRPO objective without requiring phrase-level image rewards, an additional segmentation model, or extra image generation to compute the weights. The ablations favor semantic span weighting over token-local weighting and random or fixed boundaries. Together, the simple design and supporting experiments make this a useful contribution to prompt-rewriting methods.
Critical weaknesses
  1. Preservation of the user’s intent is not sufficiently evaluated. The rewriting instructions require the model to avoid adding unrequested details, yet Figure 4 expands “four handbags” into a scene with white pedestals and a specific arrangement. Such additions may help image generation while also narrowing the user’s original request. The paper does not systematically measure unwanted additions, omitted constraints, or changes in meaning.
  2. Generalization beyond the training benchmarks remains unclear. Training uses prompts and reward models from GenEval and T2I-CompBench, and evaluation relies on the same benchmark families. The disjoint prompt splits are valuable, but the reported improvements may still depend on the particular scoring systems used during training. Evaluation with independent judges, human assessments, or a separate collection of realistic user prompts would provide stronger evidence of broader usefulness.
Questions for authors

Since training rewards and evaluation both rely on GenEval and T2I-CompBench, could you test SCOPE with an independent evaluator to check whether the improvements also hold beyond these scoring systems?

3 comments

Thank you for your careful review and constructive suggestions. We address your questions below.

Q1: How is user intent preserved? Do added details narrow the request? Are additions, omissions, and semantic changes evaluated separately?

A1: We preserve explicit requirements through rewriting instructions and image rewards evaluated against the original prompt. The system prompt in Appendix A.9 requires preserving objects, counts, attributes, text, and relations. The reward model evaluates the generated image against the original prompt.

Added details narrow an underspecified request. In the “four handbags” example, the objects and count are preserved. White pedestals add a scene element, while a fixed arrangement selects one of the layouts allowed by the original request.

We added a human audit of the rewrites for all 553 GenEval prompts, separately counting added details, omitted explicit constraints, and semantic changes. Relative to the original prompt, any detail not explicitly mentioned is counted as an “added detail,” without judging whether it is reasonable or justified. An omission means that an explicit requirement is not preserved; a semantic change means that the meaning of an original requirement is altered. Each rewrite is counted at most once in each category.

MethodPrompts evaluatedPrompts with added details (%)Prompts with omitted explicit constraints (%)Prompts with semantic changes (%)
GRPO553493 (89.15%)0 (0%)0 (0%)
SCOPE553526 (95.12%)0 (0%)0 (0%)

Q2: Are the gains recognized beyond the evaluators used during training?

A2: Yes. In the blind human evaluation, raters prefer SCOPE in every GenEval dimension in both comparisons. To test this independently of the training evaluator, we evaluate FLUX.1-dev on 553 prompts with 4 seeds per prompt, yielding 2,212 image pairs per comparison. Five raters independently assess the matched pairs in a blind evaluation.

Raters use the original prompt as the reference and make a blind forced-choice judgment with no ties. If only one image meets the original prompt’s requirements, instruction following determines the choice; otherwise, raters consider both visual quality and the degree of compliance.

GenEval dimensionImage pairsSCOPE preference over GRPOSCOPE preference over the original baseline
Single object32060.25%87.75%
Two objects39668.33%90.51%
Counting32082.00%91.56%
Colors37676.91%81.54%
Position40087.00%96.60%
Color attribution40070.00%81.80%
Overall2,21274.28%88.26%

The overall preference rates are 74.28% over GRPO and 88.26% over the original baseline, calculated by averaging across raters and weighting by the number of image pairs. Preference over GRPO reaches 82.00% for Counting and 87.00% for Position. Thus, human raters also observe the gains measured by the automatic reward models.

Thank you for the response. I will keep my positive score.

Dear Reviewer 36DZ,

Thank you very much for your continued support and positive assessment of our work. We sincerely appreciate the time and care you have devoted to reviewing our paper and considering our responses. We are grateful for your constructive suggestions and your recognition of our contribution.

Best regards,

The Authors of SCOPE

Reviewer_u58DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SCOPE, a semantic span-aware credit reweighting method for GRPO-based text-to-image prompt rewriting. It assigns shared update weights to tokens within semantic spans while keeping the image generator frozen. Experiments on GenEval and T2I-CompBench across three image generators demonstrate improvements over standard GRPO and the reported baselines.

Critical strengths

The method is relatively simple and compatible with existing GRPO training pipelines. The experiments show consistent improvements across multiple image generation backbones and compositional benchmarks.

Critical weaknesses
  1. Unjustified details introduced during prompt rewriting.

Many qualitative examples primarily add details that cannot be inferred from the original prompts, such as specific spatial arrangements, backgrounds, lighting, and camera settings. These additions may improve benchmark scores but do not necessarily reflect the user's original intent. The paper lacks a systematic evaluation of semantic preservation or unintended content introduction.

  1. Insufficient comparison with existing prompt engineering approaches.

Although the paper includes GRPO, RePrompt, and a training-free rewriter, the comparisons with strong conventional LLM-based rewriting and prompt engineering methods remain limited. In particular, some RePrompt results are taken from prior work rather than reproduced under fully matched conditions. It is therefore difficult to assess the practical advantage of SCOPE over established prompt optimization approaches.

Questions for authors

How do you evaluate whether rewritten prompts preserve the original user intent rather than introducing unsupported details? Are image alignment scores evaluated against the original prompts or the rewritten prompts?

How does SCOPE compare with strong LLM-based prompt rewriting and prompt engineering baselines under matched evaluation settings?

1 comment

Thank you for your careful review and constructive suggestions. We address your questions below.

Q1: How do you check whether rewrites preserve user intent without unsupported additions?

A1: We preserve user intent through both rewriting instructions and image rewards evaluated against the original prompt. The system prompt in Appendix A.9 requires preserving objects, counts, attributes, text, and relations. The reward model evaluates the generated image against the original prompt.

Prompt rewriting can turn a short request into a concrete visual description by adding layout, background, and other details. We therefore separately inspect the introduction of additional details and whether explicit requirements are omitted or changed. We added a human audit of 553 GenEval prompt rewrites on Qwen-Image, separately counting added details, omitted explicit constraints, and semantic changes. Relative to the original prompt, any detail not explicitly mentioned is counted as an “added detail,” without judging whether it is reasonable or justified. An omission means that an explicit requirement is not preserved; a semantic change means that the meaning of an original requirement is altered. Each rewrite is counted at most once in each category.

The table reports the number and percentage of prompts in each category.

MethodPrompts evaluatedAdded detailsOmitted constraintsSemantic changes
GRPO553493 (89.15%)0 (0%)0 (0%)
SCOPE553526 (95.12%)0 (0%)0 (0%)

Q1.a: Are image-alignment scores computed from the original prompt or the rewritten text?

A1.a: Image-alignment scores are always computed from the original prompt; the rewritten text is used only to generate the image. This keeps the optimization target aligned with the requirements explicitly stated by the user.

Q2: How does SCOPE compare with stronger LLM rewriters under matched evaluation settings?

A2: SCOPE achieves a GenEval Overall score of 0.95, while the tested LLM rewriters and the released PromptEnhancer model score between 0.88 and 0.89. All methods use the same Qwen-Image generator, GenEval prompts, image-generation settings, and sampling seeds for this comparison.

MethodQwen-Image GenEval Overall ↑
Original prompt0.87
Officially released PromptEnhancer rewriter0.88
GPT 6 Astra: training-free rewriting0.88
Claude Fable 5.2: training-free rewriting0.88
Gemini 4 Carbon: training-free rewriting0.89
GRPO0.91
SCOPE0.95

GPT 6 Astra, Claude Fable 5.2, and Gemini 4 Carbon use a shared rewriting instruction adapted from Appendix A.9. The instruction requires preserving objects, counts, attributes, text, and relations, clarifying visual details supported by the original text, avoiding unsupported additions, and producing a concise plain-text prompt without span delimiters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.