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

The gist The SCOPE framework introduces Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation, which aligns update weighting with coherent semantic spans to resolve subword credit fragmentation in reinforcement learning prompt rewriting.

How it works

1. SCOPE is an objective based on Group Relative Policy Optimization (GRPO) that aligns update weighting with the semantic structure of compositional edits, as described on page 2.

2. The rewriter jointly generates an enhanced prompt and its semantic partition, and SCOPE pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses while retaining a single advantage for each completion, according to page 4.

3. This coefficient modulates all token updates within the span, as shown in Figure 2 on page 4.

4. The objective retains a loss for each valid token and a single advantage for each completion, without separate span rewards or additional image renders or a learned critic, as stated on page 5.

Key Mechanisms

** Grouped Rewards and Advantages**

The policy maximizes the expected reward over prompts drawn from the input distribution D and completions sampled from πθ, as formulated in equation (1) on page 3. For each minibatch prompt xg, N-1 rewrites are sampled, and a reference prompt is appended to contribute to advantage estimation but receive no policy update, according to page 3. The reward combines image alignment and output format: rg,i = wt2i rt2i(xg, c(yg,i)) + wfmt rfmt(yg,i), as shown in equation (2) on page 3.

** Semantic Span Credit Modulation**

The rewriter generates the rewrite and its boundaries in one completion, placing only between adjacent nonempty spans, and tokenizer character offsets link the extracted spans to the positions in the loss, as detailed on page 3. SCOPE measures the log-probability shift on the sampled completion after each update by calculating ∆g,i,t = log πθ(yg,i,t xg, yg,i,<t) − log πold(yg,i,t xg, yg,i,<t), as shown in equation (6) on page 3. Averaging the log ratios ∆g,i,t within each span removes the direct scaling with token count to derive bg,i,s = 1/Tg i, s X t∈Tg i, s ∆g i,t, as shown in equation (7) on page 3. The exponential of bg,i,s is the geometric mean of the ratios ρg i,t within span s. The sensitivity parameter β controls the response to this aggregated signal with clipping limiting the raw coefficient: vg i,s = clip (exp(βbg i,s), 0.8, 1.2), as shown in equation (8) on page 3. The normalized coefficient is calculated as wg i,s = P vg i,s / mean(v) Tg i, rvg i, r, as shown in equation (9) on page 3.

Results and Significance

** Performance Gains**

Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE consistently outperforms vanilla GRPO on GenEval [13] and T2I-CompBench [18], achieving average relative gains over vanilla GRPO of 9.48% on GenEval and 7.62% on T2I-CompBench, as reported on page 5. The ablations favor pooling policy shifts within semantic spans over weighting individual tokens, and semantic boundaries over random boundaries or fixed windows, supporting semantic structure as a practical inductive bias for prompt engineering with RL, according to page 5.

** Mechanistic Insights**

A crucial finding is that applying likelihood shifts at individual tokens degrades performance below standard GRPO across all backbones, confirming the hypothesis that unpooled token-level modulation injects high-variance subword noise that disrupts coherent multi-token phrases, whereas pooling shifts within semantic spans stabilizes credit allocation, as detailed on page 8. Furthermore, replacing semantic partitions with random boundaries or fixed 3-token windows causes a severe 8–9% drop relative to SCOPE on Qwen-Image, proving that linguistic boundary alignment is essential for effective credit assignment, according to page 8.

** Qualitative Superiority**

SCOPE lifts a frozen open-weight Qwen-Image to the level of the commercial gpt-image 2 on layout and counting by learning to keep each constraint as an intact phrase rather than scattering it across subword tokens, as shown in Figure 4 on page 9. On counterfactual attributes, SCOPE goes further than other methods because it is only model that renders a white orange, while gpt-image 2 falls back to an orange-colored prior, according to page 9.

SCOPE establishes structure-aligned credit assignment as a plug-and-play paradigm for unlocking the compositional potential of frozen visual generators across Qwen-Image, SD3, and FLUX.1-dev by pooling policy log probability shifts within coherent semantic spans to modulate a shared completion advantage, as concluded on page 10.

--- Page 1

SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation</ref:2610.22264

--- Page 3

Compositional prompt edits to objects, attributes, counts, spatial relations, and scene context form coherent phrases that can span several subword tokens.</ref:5

--- Page 4

Figure 1. Comparison of uniform token weighting in GRPO and semantic span weighting in SCOPE.</ref:8

--- Page 5

composition holistically, without direct attribution to constituent spans or tokens. This exposes a fundamental credit assignment bottleneck: uniform token-level updates fragment coherent compositional edits across arbitrary subword boundaries, yet estimating dense per-span rewards incurs prohibitive rendering overhead.</ref:10

--- Page 5

We introduce SCOPE, an objective compatible with GRPO that pools policy log probability shifts within semantic spans to assign shared token weights.</ref:2

--- Page 5

The rewriter outputs a prompt rewrite and its semantic partition, mapped onto rollout tokens.</ref:3

--- Page 6

The alignment term rt2i ∈ [0, 1] uses the GenEval and T2I-CompBench evaluation models to score the cleaned prompt and rendered image.</ref:11

--- Page 7

For completion yg,i = (yg,i,1,..., yg,i,T), let Vg i contain its valid token positions.</ref:3.2

--- Page 8

GRPO (uniform token level) SCOPE span credit same advantage copied to tokenizer cells span pooled influence modulates the same advantage</ref:4.1

--- Page 9

The normalized coefficient is detached and applied to the advantage at each position in its span: Aeg i,t = Ag i stopgrad(wg i,s), t ∈ Tg i,s</ref:5

--- Page 10

Lpol = 1/B X g∈B (N − 1)−1 P i<N lscope g,i</ref:12

--- Page 6

Effect of credit weighting. Figure 3(a) compares SCOPE’s span-pooled weighting against uniform GRPO and unpooled token-level likelihood shifts.</ref:4.3

--- Page 8

unpooled token-level modulation injects high-variance subword noise that disrupts coherent multi-token phrases, whereas pooling shifts within semantic spans stabilizes credit allocation.</ref:7

--- Page 9

--- Page 10

The positive coefficient preserves the sign of the completion advantage and scales the token loss contributions within that span.</ref:5

--- Page 10

We demonstrate that SCOPE consistently outperforms vanilla GRPO on GenEval [13] and T2I-CompBench [18]. These gains hold across three frozen T2I backbones: Qwen-Image, SD3, and FLUX.1-dev.</ref:13

--- Page 6

Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE consistently outperforms vanilla GRPO on GenEval [13] and T2I-CompBench [18].</ref:4.2

--- Page 7

SCOPE exceeds GRPO in all six reported categories on each frozen backbone (Table 2).</ref:6

--- Page 7

Average scores improve by 6.7%, 8.0%, and 8.2% on Qwen-Image, SD3, and FLUX.1-dev, respectively.</ref:4.2

--- Page 9

SCOPE achieves the highest overall GenEval score across all three frozen backbones (Table 1), boosting overall performance over vanilla GRPO from 0.91 to 0.95 on Qwen-Image (+4.4% relative), 0.

1. Bold header: Semantic credit assignment for compositional edits. SCOPE pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses while retaining a single advantage for each completion. This resolves credit fragmentation by ensuring compositional prompt edits form coherent semantic spans whose boundaries need not align with the subword boundaries imposed by the tokenizer.

2. Bold header: Superior performance on complex compositional tasks. SCOPE consistently outperforms vanilla GRPO on GenEval and T2I-CompBench, achieving gains like +4.4% relative on Qwen-Image and substantial improvements in object counting and spatial positioning.

3. Bold header: Robustness against credit assignment noise. The method shows that unpooled token-level modulation injects high-variance subword noise that disrupts coherent multi-token phrases, while SCOPE's strategy stabilizes credit allocation by using span pooling, even when compared to random boundaries or fixed windows.

4. Bold header: Fine-grained control over visual attributes. The system allows for precise manipulation of specific constraints, as demonstrated by the rewriter learning to keep each constraint as an intact phrase rather than scattering it across subword tokens, enabling it to render counterfactual attributes like a white orange instead of an orange prior.

5. Bold header: Scalable and efficient deployment. SCOPE operates by optimizing the policy without requiring separate span rewards, additional image renders, or a learned critic, allowing for seamless integration into existing T2I reward pipelines across various frozen backbones like Qwen-Image, SD3, and FLUX.1-dev.

Reinforcement learning (RL) enables language models to improve text-to-image (T2I) generation by rewriting prompts for a frozen image generator. Existing methods typically apply a single signal derived from the image to individual subword tokens, although objects, counts, attributes, and relations form coherent semantic spans. We propose SCOPE, an objective based on Group Relative Policy Optimization (GRPO) that aligns update weighting with this structure. The rewriter jointly generates an enhanced prompt and its semantic partition. SCOPE pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses while retaining a single advantage for each completion. It requires neither separate span rewards nor additional image renders. Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE achieves average relative gains over vanilla GRPO of 9.48% on GeneVal and 7.62% on T2I-CompBench. Ablation studies show that both semantic boundaries and pooling within spans contribute to these gains, supporting semantic structure as a practical inductive bias for prompt engineering with RL.