- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies failures of text-to-image models when a prompt implies visual content through idioms, references, or physical reasoning rather than stating it explicitly. It proposes bridgeprompt, which uses GEPA to optimize a reusable instruction for a frozen language-model rewriter using image-level feedback from a frozen VLM. Experiments on T2I-ReasonBench and transfer settings evaluate whether the learned instruction improves reasoning fidelity without modifying the generator.
Critical strengths
-
The paper addresses a real problem in text-to-image generation. T2I models sometimes struggle with prompts that require interpreting implicit meanings, resolving indirect references, or understanding physical and scientific relationships.
-
The improvements on T2I-ReasonBench are substantial. BridgePrompt raises the score of SD3.5 Medium from 46.8 to 74.6, a gain of 27.8 points. The improvements are consistent across all four categories, with particularly strong gains on idiom and entity reasoning.
Critical weaknesses
Thank you for the additional experiments and clarifications. The cross-generator and transfer results partially address my concerns about scope and relevance, but the central methodological concern (W1) remains.
What is addressed. The Qwen-Image experiments are the most convincing new evidence for W3. Generator-specific recompilation reaches 80.5 (+13.0 over raw prompting) on a modern model with an LLM text encoder, which shows the formulation remains useful beyond SD3.5 Medium. I appreciate the honest reporting that direct transfer barely exceeds GPT-4o zero-shot refinement (73.5 vs 73.1) and degrades textual-image-design (77.1 → 67.9); this tempers the "reusable instruction" claim and is the right framing. For W2, the R2I-Bench (0.47 → 0.63) and WISE (0.45 → 0.71) transfer results are a reasonable test of distribution shift. I look forward to the released prompt IDs.
What remains unaddressed (W1). The response restates the formulation but does not provide the evidence I asked for. The central claim is that image-level feedback is the key ingredient, yet there is no ablation isolating image-level from text-level feedback — a comparison the paper itself flags as missing (§5.1 / Table 8). Two further issues compound this:
- The instruction guard (Table 6) carries most of the optimization gain. Without it, the score drops to ~55.4, only 1.5 points above the seed instruction (53.9). The bulk of the 53.9 → 74.6 improvement therefore depends on a hand-written engineering filter rather than the optimized instruction text. The rebuttal does not discuss this.
- There is still no comparison with prompt-refinement methods that train or iterate a rewriter (RePrompt, Promptist, OPT2I). The relevant question for a "frozen-prompter" claim is whether a trained rewriter closes the same gap more cheaply.
The GPT-4o-authored instruction (59.6 vs 74.6) is a useful new baseline and supports that optimization matters, but it does not isolate what is being optimized toward — image feedback versus instruction search.
Assessment. The problem is real, the gains are substantial, and the transfer and cross-generator evidence is stronger than I expected going into the rebuttal. However, the methodological contribution is still primarily an application of GEPA to T2I reinforced by a hand-written guard, and the single ablation that would establish image-level feedback as load-bearing is absent. I maintain my score as 2.
Questions for authors
No further questions
Response to Reviewer 5kNe
We thank the reviewer for recognizing the importance of implicit visual reasoning and the improvements across all four T2I-ReasonBench categories. We address the concerns about novelty, generalization, and newer generators below.
W1 / Q1: Technical contribution beyond applying GEPA
BridgePrompt addresses the implicit visual meaning gap through instruction-level optimization. Many prompts convey an intended meaning without explicitly stating the visual relationships needed for generation. We treat this as a conditioning-interface problem and optimize the instruction guiding a frozen rewriter to make those relationships explicit.
Unlike instance-specific prompt optimization, BridgePrompt learns one instruction shared across inputs, using feedback from generated images. Both the prompter and generator remain frozen, and inference requires only one rewrite and one generation. The optimization cost is thus amortized across unseen prompts.
The results support this formulation:
The 20.7-point gain over the initial instruction demonstrates the value of optimization beyond simply providing a plausible instruction. Best-of-50 sampling, CFG search, and length-control diagnostics further indicate that the gains cannot be explained solely by increased sampling or verbosity.
Our contribution is not a new GEPA algorithm, but a reusable, image-grounded optimization formulation that improves how implicit visual meaning is communicated to a frozen generator.
W2 / Q2: Generalization beyond T2I-ReasonBench
We evaluate the learned instruction on two additional benchmarks without reoptimization. On R2I-Bench, the Commonsense Reasoning (695 prompts) and Causal Reasoning (151 prompts) subsets show an improvement from 0.47 to 0.63. On WISE, the Cultural Commonsense (400 prompts) and Natural Science (300 prompts) subsets improve from 0.45 to 0.71.
These evaluations follow the respective official protocols and test distinct prompt distributions. The results demonstrate that the instruction transfers beyond the categories used during optimization. We will release the evaluated prompt IDs with our code for reproducibility.
W3 / Q3: Relevance to contemporary image generators
Our evaluation also includes SDXL, Sana-1.5, FLUX.1-dev, and Qwen-Image, extending beyond SD3.5 Medium.
On Qwen-Image, raw prompting scores 67.5, GPT-4o refinement 73.1, and direct BridgePrompt transfer 73.5. Optimizing the instruction specifically for Qwen-Image raises performance to 80.5, a 13.0-point gain over raw prompting.
These results show that stronger generators still benefit from making implicit visual constraints explicit. They also distinguish direct transfer from generator-specific optimization, which achieves larger gains.
Direct transfer does reduce textual-image-design performance from 77.1 to 67.9, indicating that instruction effectiveness depends on the target generator. Nevertheless, the aggregate improvements and stronger generator-specific result support the relevance of our formulation beyond SD3.5 Medium.
The qualitative examples highlight the failure mode. The quantitative cross-generator results, particularly the Qwen-Image evaluation, provide the stronger evidence that the proposed formulation remains useful beyond the main SD3.5 Medium setting.