- Rating
- 3
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This work replaces the open-ended textual reasoning that normally precedes UMM image generation with VSRL, a rigid structured representation covering the global scene, object IDs and attributes, plus their pairwise relations. The authors build triplets of prompt, VSRL and corresponding image, then supervise-finetune the UMM to output VSRL and draw images conditioned on it. They further train with StruFlow-GRPO, using programmatic structure rewards alongside image-level rewards. On the BAGEL benchmark suite, VSRL outperforms both the base model and vanilla Self CoT across five generation benchmarks; for instance, GenEval Short jumps from 79.0 to 88.9, and OneIG rises from 32.4 to 43.7. On Cheers, they train VSRL and regular natural-language CoT on identical underlying facts using the exact same training setup, and VSRL comes out ahead consistently on GenEval, TIIF and DPG.
Critical strengths
The Cheers control experiment is well conceived. It uses identical prompts, target images, initialization and training schedule; the natural-language chain-of-thought text is simply transcribed from the same factual content used for VSRL. This makes the performance gap difficult to write off as just better training data, and that is exactly why the representation argument lands for me.
Object IDs together with explicit relation anchors offer a practical fix for common CoT breakdowns: instances from the same category getting merged together and spatial descriptions drifting. The improvements seen on GenEval Short and benchmarks heavy on counting and relational reasoning align directly with this design choice, rather than being a generic boost from reinforcement learning.
Critical weaknesses
Credit assignment on BAGEL is unclear. The baseline Self CoT already went through its own training pipeline, while VSRL adds fresh SFT data plus StruFlow-GRPO on top. I would like to see ablations at matched training steps: VSRL with only SFT, then SFT plus image reward, then SFT plus both image and structure rewards. Without this breakdown, Table 1 risks overstating gains from the representation itself when they may simply come from extra optimization.
There is also very little evaluation of VSRL outputs at inference time. We do not know how frequently the model lists an object that never appears in the final image, or how often it states spatial relations backwards. If the structure reward only pushes for valid, complete parses, the model could produce clean VSRL schemas that the image generator still ignores. A small consistency check across prompt, VSRL and generated image would make the whole mechanism far more credible.
Questions for authors
We request ablations on a single backbone with identical update counts: VSRL under SFT alone, SFT plus only the image reward, and the full StruFlow-GRPO setup.
We also want breakdowns for three failure modes on test cases: (a) VSRL fails to include an object specified in the prompt, (b) the final image leaves out an object listed in VSRL, and (c) stated relations do not match the image. We need to see whether RL mainly cuts down on (b) and (c), or mostly cleans up (a).
For the Cheers NL-CoT control, please verify that its reasoning template does not inject bounding boxes or coordinate information unavailable to VSRL, and that the maximum reasoning token budget and image decoding parameters are matched across both conditions.
4 comments
Q2/W2: Prompt–VSRL–Image Consistency Analysis and Q3: Information Consistency and Evaluation Settings for the Cheers Natural-Language Conversion
Q2/W2: Prompt–VSRL–Image Consistency Analysis
We appreciate your suggestion to evaluate VSRL reasoning quality, as this evaluation is important for assessing how effectively VSRL guides image generation. Therefore, following your suggestion, we conducted an additional Prompt–VSRL–Image consistency analysis.
We compared VSRL-SFT and VSRL-StruFlow-GRPO on GenEval Short and DPG, measuring the following three types of errors:
- (a) Prompt-to-VSRL object omission: objects required by the prompt but omitted from VSRL.
- (b) VSRL-to-image object omission: objects specified in VSRL but missing from the generated image.
- (c) VSRL-to-image relation mismatch: relations specified in VSRL but not correctly realized in the generated image.
In this evaluation, each sample consists of the original prompt, the VSRL generated by the model, and the final generated image. We use Qwen3.8-27B as the judge and measure these three types of errors at the level of individual Prompt–VSRL–Image rollouts.
The results are shown below. Parentheses indicate relative reductions compared with SFT.
| Method | GenEval a ↓ | GenEval b ↓ | GenEval c ↓ | DPG a ↓ | DPG b ↓ | DPG c ↓ |
|---|---|---|---|---|---|---|
| VSRL-SFT | 0.0000% | 0.3948% | 3.5035% | 2.3308% | 1.8952% | 9.1005% |
| VSRL-StruFlow-GRPO | 0.0000% (unchanged) | 0.3130% (↓20.73%) | 3.4481% (↓1.58%) | 2.2627% (↓2.92%) | 1.5292% (↓19.31%) | 8.9957% (↓1.15%) |
The results show that the improvements from RL are not limited to object coverage in VSRL relative to the prompt (a); we also observe reductions in the rates of missing objects in the generated images (b) and relation mismatches (c), the two issues you specifically raised. These results indicate that StruFlow-GRPO not only helps VSRL describe the objects required by the prompt more completely, but also improves how well the final image matches the objects and relations specified in VSRL, thereby improving VSRL's ability to guide image generation.
We will add a new subsection, “Prompt–VSRL–Image Consistency Analysis,” to Appendix A of the revised manuscript, including the full results, evaluation procedure, and statistical details.
Q3: Information Consistency and Evaluation Settings for the Cheers Natural-Language Conversion
We apologize for not explaining these experimental details clearly enough in the original manuscript and for the resulting confusion.
First, we confirm that the natural-language CoT conversion templates for Cheers do not add bounding boxes or coordinates that are not present in the original VSRL. As shown in the original Appendix C.1, “Control Construction,” and C.2, “Paired Supervision Example,” we use fixed templates and connecting words to convert the objects, attributes, and relations in VSRL into natural language, while preserving the object references at both endpoints of each relation and adding no visual information absent from the original VSRL.
Second, in the Cheers comparison between VSRL and natural-language CoT, all inference settings used in the evaluation are kept the same, including the maximum number of reasoning tokens and the image decoding settings you mentioned. We apologize for not making this clear in the original manuscript. We will clarify this in the existing Appendix D, “Supervised Fine-Tuning and Inference Configurations,” and explicitly report the corresponding settings.
Thank you again for your careful review. Your suggestions have not only helped us add the necessary ablation and consistency experiments, but also encouraged us to explain more clearly how VSRL is optimized and how it affects final image generation. This has helped us improve the paper.
Great, I tend to accept this paper due to authors meaningful feedback and solution.
Thank you for your encouraging feedback and positive assessment of our work. We are delighted that our additional experiments and clarifications have helped address your concerns, and we are truly encouraged by your support.Your constructive suggestions have also helped us strengthen the experimental evidence and clarify our method and experimental setup. We will incorporate these additional results and details into the revised manuscript.
Q1/W1: RL Reward Ablation
Reviewer 3 (a959)
We appreciate your detailed and constructive questions and your help in identifying gaps in the experimental evaluation in the original manuscript. Following your suggestions, we conducted an additional RL reward ablation across multiple benchmarks, as well as a Prompt–VSRL–Image consistency analysis. Your suggestions have helped us better understand and evaluate the method. We respond to your comments below.
Q1/W1: RL Reward Ablation
Thank you for your suggestion regarding the attribution of performance gains on BAGEL. We agree that clarifying the contributions of different optimization factors helps us better assess VSRL's effectiveness.
Therefore, we conducted an additional 200-step RL ablation using only image rewards on BAGEL. The image-reward-only variant and full StruFlow-GRPO start from the same VSRL-SFT model and use the same trainable modules, number of training steps, and other RL settings, differing only in the reward components. Both RL variants are trained for 200 steps, while VSRL-SFT serves as their common initialization without additional RL updates.
We compared the methods on four generation benchmarks with five evaluation settings: GenEval, DPG, TIIF Short/Long, and OneIG. The results are as follows:
We introduced structure rewards so that optimization considers both the quality of the final generated image and the structural quality of the VSRL generated by the model. Specifically, structure rewards evaluate aspects such as VSRL's structural validity, coverage of object categories required by the prompt, and structural richness. Together with image rewards, they contribute to trajectory returns and model parameter updates, allowing the model to use feedback on intermediate VSRL to optimize toward structures better suited to guiding image generation.
The results show that, under the same initialization and training budget, full StruFlow-GRPO outperforms image-reward-only RL in all five evaluation settings. This further indicates that the structural quality of VSRL is an important factor affecting final image generation, and that including it in optimization can lead to better generation performance.
We will add this RL reward ablation and its settings to the existing Appendix A.2, “Training Stages and Optimization Scope,” in the revised manuscript.