Pith. sign in

REVIEW 3 major objections 6 minor 63 references

RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A trained prompting policy uses chain-of-thought reasoning and image-level rewards to fix spatial and compositional failures in text-to-image generation, lifting GenEval position accuracy from 0.18 to 0.62 on FLUX and overall accuracy to…

desk verdict Sensible reprompting method with impressive GenEval gains, but the missing train/test separation details and a flawed variance-reduction proof keep me from trusting the headline numbers. read the letter →

arxiv 2505.17540 v1 pith:Z334Q7I3 submitted 2025-05-23 cs.CV cs.AI

Mingrui Wu , Lu Wang , Pu Zhao , Fangkai Yang , Jianjin Zhang , Jianfeng Liu , Yuefeng Zhan , Weihao Han
show 9 more authors
Hao Sun Jiayi Ji Xiaoshuai Sun Qingwei Lin Weiwei Deng Dongmei Zhang Feng Sun Qi Zhang Rongrong Ji
This is my paper
classification cs.CVcs.AI
keywords text-to-imagegenerationpromptenhancementreinforcementlearningchain-of-thoughtreasoningGRPOcompositionalfidelityspatialrewardmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

A short, under-specified prompt such as 'a photo of a couch below a vase' trips up text-to-image models, which often render physically implausible scenes. The paper argues that the fix is not a bigger image model or handcrafted prompt rules, but a reprompting policy trained to reason about the scene before writing the final prompt. RePrompt trains a 3-billion-parameter language model by reinforcement learning, using image-level rewards for human preference, semantic alignment, and visual composition, so the policy learns to resolve spatial and object relations instead of merely elaborating style. The central result is that this trained policy lifts GenEval position accuracy from 0.18 to 0.62 on FLUX and pushes overall accuracy to 0.76, while transferring to other text-to-image backbones at lower latency than iterative refinement. If right, it suggests that reasoning about the visual implications of a prompt, learned from downstream image feedback, can close part of the semantic-visual gap without retraining the image generator.

What carries the argument

The load-bearing mechanism is the prompt policy's structured output: a chain-of-thought reasoning trace followed by an enhanced prompt, trained as a single-step Markov decision process with the text-to-image model kept frozen. The policy is updated by Group Relative Policy Optimization, where each sampled reasoning-and-prompt pair is rendered into an image and scored by an ensemble reward combining human-preference feedback, vision-language semantic alignment, output-format compliance, and a length constraint. Because the reward depends only on the prompt-image pair and not on the image model's internals, the same learned policy can be applied to any frozen diffusion backbone.

What would settle it

Generate a held-out evaluation set from the same six templates but with objects and spatial relations not present in the training list, and also a paraphrased set that breaks template phraseology; if position scores drop toward the baseline on both, the 0.62 gain is largely template memorization rather than transferable spatial reasoning.

Watch

Extended reading notes

Core claim

RePrompt trains a 3-billion-parameter language model (Qwen2.5-3B) to output a self-reflective reasoning trace followed by an enriched prompt, optimizing the pair with Group Relative Policy Optimization against an ensemble reward computed from images rendered by a frozen text-to-image model. On GenEval, the policy raises FLUX's position score from 0.18 to 0.62, improves counting by 22.2 percent, and lifts overall accuracy to 0.76, with consistent gains when the same trained policy is applied to SD3 and PixArt-Σ. The improvement transfers to T2I-Compbench, especially spatial composition, and takes about 30 seconds per image versus 140 seconds for an iterative baseline. The authors frame this as evidence that explicit, reward-grounded reasoning during prompt construction can substitute for expensive inference-time optimization and for larger prompt-enhancement models.

Load-bearing premise

The load-bearing premise is that the 9,000-prompt training corpus, assembled from six object-centric templates and a GPT-4-generated object list, is truly disjoint from GenEval's evaluation prompts, so the reported gains reflect compositional generalization rather than training and evaluation similarity.

Editorial extensions

If this is right

  • Trained once on FLUX, the same reprompting policy improves GenEval overall accuracy on SD3 from 0.69 to 0.75 and on PixArt-Σ from 0.54 to 0.62, so the learned reasoning style is not tied to one image model.
  • On the GenEval position category, the policy raises relative scores by 77.1 percent on FLUX, 78.8 percent on SD3, and 122.2 percent on PixArt-Σ over the Qwen2.5-3B enhanced baseline.
  • The single-pass design reaches 0.76 GenEval accuracy in about 30 seconds per image, whereas the Idea2Img baseline needs roughly 140 seconds to reach 0.69, indicating that reward-driven reasoning can replace expensive iterative self-correction.
  • Ablations show that supervised fine-tuning and reinforcement learning contribute differently: SFT injects object-attribute priors, while RL drives the large spatial and counting gains, with the combination yielding the best overall result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to vary the reasoning trace's granularity, for example forcing the model to enumerate every object and its spatial predicate, and measuring whether reward variance falls and position accuracy rises further, as the paper's variance-reduction argument predicts.
  • Because the reward is backbone-agnostic, the same pipeline could be pointed at the harder compositional categories, counting and attribute binding, where the reported gains are smaller, by adding reward terms specialized to those failures.
  • If the training corpus closely mirrors GenEval's construction, the position gain may partly measure how well a prompting policy exploits the benchmark's template grammar; a domain-shift evaluation on paraphrased or natural-language prompts would separate genuine composition learning from template mimicry.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes RePrompt, a framework that fine-tunes a small language model (Qwen2.5-3B) to generate a structured reasoning trace followed by an enhanced prompt for text-to-image generation. The policy is trained with supervised fine-tuning followed by group-relative policy optimization (GRPO), using a composite reward that combines ImageReward, VLM-Reward, a syntax-format reward, and a token-length reward, with FLUX.1-dev as the fixed image synthesizer. The trained reprompting policy is evaluated on GenEval and T2I-Compbench across three backbones (FLUX, SD3, PixArt-Sigma), reporting substantial gains in spatial position (e.g., 0.35 to 0.62 on FLUX GenEval Position) and overall accuracy (0.76 on GenEval), plus ablations of SFT, RL, reasoning, and reward components.

Significance. If the claimed generalization holds, RePrompt is a practically valuable contribution: it decouples prompt generation from image generation, trains without human-annotated reasoning traces, and uses external reward models, so the main result is not definitionally circular. The paper provides a code link, its ablations isolate the contributions of SFT, RL, reasoning, and each reward component, and the cross-backbone evaluation is a useful stress test of model-agnostic reprompting. However, the central claim of compositional generalization rests on an unspecified train/test overlap filter for the training corpus, and the theoretical variance-reduction argument in Appendix A is mathematically incorrect as written. These issues, together with the absence of error bars, currently prevent the paper from fully supporting its state-of-the-art claim.

major comments (3)
  1. [§4.1 (Training Data)] The training corpus is constructed by adapting GenEval's six object-centric templates to a GPT-4-generated list of 288 objects, and the paper states it was 'carefully filtered to avoid overlap with the GenEval' without specifying the filtering rule. Because GenEval's test prompts are themselves generated from the same six template families, exact-string deduplication would leave the template-level distributions of training and test nearly identical, so the headline Position improvement (0.18 to 0.62 on FLUX in Table 1) and the Counting/Attribute gains could reflect train/test distribution similarity rather than compositional generalization to unseen prompt structures. Please specify the overlap-removal procedure, release the training prompts, and add evaluations on held-out template families or object-relation combinations that are disjoint from training (e.g., novel objects and unseen relation types); without this, the central claim of compositional generalization is not established.
  2. [Appendix A (Theorem A.1)] Theorem A.1 states Var(r(H,P')) = E_H[Var(r|H)] = Var(r(P')), but by the law of total variance, Var(r(H,P')) = E_H[Var(r|H)] + Var_H(E[r|H]), so the claimed equality omits the between-H term and the inequality as written does not follow. Since Section 3.3 invokes this variance-reduction analysis as a justification for faster and more stable GRPO training, the theoretical claim is load-bearing. Please correct the identity (e.g., state that conditioning on H reduces the conditional variance of the per-group estimator, if that is the intended claim) or remove the proof and rely on the empirical ablations.
  3. [Tables 1-4 and Appendix C] All quantitative results are reported without error bars or multiple seeds; ablations such as Table 3 (SFT-only improvement of +0.01 overall) and Table 7 (reward ablations) involve differences smaller than typical run-to-run variance for RL-trained policies. Please report mean and standard deviation over at least three training runs for the main tables, or otherwise quantify the stability of the reported gains.
minor comments (6)
  1. [Figures 1, 5, 8-11] The model name 'DELL-E3' is a typo throughout the figures and should be 'DALL-E 3'.
  2. [Table 7] The row labeled 'R1-Prompter' with alpha=1, gamma=0 is actually RePrompt without VLM-Reward, not a separate baseline; rename the row to avoid confusion with the related-work method T2I-R1.
  3. [Appendix D] The text refers to 'Figure 7' and 'Figure 9' when discussing position and two-object examples, but the corresponding figures are numbered 8 and 10; renumber the cross-references.
  4. [Table 4] Table 4 reports accuracy on 'the subset of Geneval' but compares with Table 1's overall GenEval accuracy, and the same value (0.76) appears in both; clarify whether the subset is the full GenEval set or a different split.
  5. [§4.1] The text says the VLM used for computing VLM-Reward is GPT-4V; specify the exact model version and decoding settings for reproducibility.
  6. [§4.1 (Training Data)] Please provide several example prompts from the 9,000-prompt training corpus so that readers can independently assess the overlap risk with GenEval.

Circularity Check

1 steps flagged · score 2.0 of 10

Central RL pipeline is independently benchmarked and not circular; only the Appendix A variance-reduction 'proof' is true by definition and is non-load-bearing.

  1. self definitional [Appendix A.2, Theorem A.1 proof (also cited in Sec. 3.3)]
    "By definition, Var(r(H, P′)) = EH [Var[r | H]] and since Var[r(P′)] = EH [Var[r | H]] + VarH [E[r | H]], the nonnegativity of VarH [E[r | H]] yields the result."

    The proof defines the reasoning-augmented variance Var(r(H,P')) to be exactly the within-H conditional variance EH[Var[r|H]], while expanding the bare-prompt variance Var(r(P')) to include the additional nonnegative between-H term VarH[E[r|H]]. The claimed inequality Var(r(H,P')) ≤ Var(r(P')) is therefore true by construction, not by any property of the reasoning trace H. The theorem's conclusion is baked into its 'by definition' line, so no independent variance-reduction mechanism is established. This is a supporting theoretical claim rather than the source of the empirical benchmark results.

full rationale

By the paper's own equations, the central training loop is not circular: the policy is optimized via GRPO against Rtotal = Rvis + Rstruc + Rlen, where Rvis uses the external ImageReward and VLM-Reward models, and the reported GenEval and T2I-Compbench numbers come from fixed benchmark evaluators rather than from the training objective. No load-bearing assumption is imported from the authors' own prior work; citations to GRPO/DeepSeek-R1, ImageReward, GenEval, and similar works are external and independently checkable. The only step that reduces to its own definition is Appendix A's variance-reduction 'proof,' which defines Var(r(H,P')) as the conditional variance EH Var[r|H] and compares it with a bare variance that includes the between-H term. That makes the inequality true by construction, but it is a supporting theoretical footnote, not the source of the empirical gains. The Section 4.1 training corpus is built from GenEval-style templates and only described as 'carefully filtered'; this is a potential distribution-overlap confound for the generalization claim, but it is not circularity because GenEval accuracy is not used as a reward or fitted parameter. Overall, the central derivation is self-contained and independently evaluated; score 2 reflects the one non-load-bearing self-definitional proof.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The policy is trained with an ensemble reward whose weights and length bounds are hand-set, and it relies on external reward models and a template-generated training distribution. No new physical entities or fitted constants beyond standard hyperparameters are introduced; the main unverified premises are the reward models' validity and the train/eval distribution separation.

free parameters (3)
  • ImageReward weight α = 0.5
    Set equal to VLM-Reward weight in the ensemble reward (Eq. 1, Table 5); no sweep reported.
  • VLM-Reward weight γ = 0.5
    Set equal to ImageReward weight; no sweep reported.
  • Prompt length bounds Lmin, Lmax = 15, 77 tokens
    Empirically chosen to match T2I model token limits (Eq. 3, Table 5).
assumptions (4)
  • domain assumption ImageReward and VLM-Reward are valid proxies for human preference and semantic alignment of generated images.
    The entire RL objective Rvis relies on these external reward models (Eq. 1); if they misjudge image quality, the learned policy is misaligned.
  • domain assumption The fixed T2I model f_phi can be treated as a stochastic black box; reward depends only on the generated image and text.
    The MDP formulation and GRPO update assume no gradient through f_phi and that Rtotal(I) is a sufficient training signal (Section 3.1).
  • standard math GRPO with group size 4 gives stable and unbiased policy updates for this single-step MDP.
    The clipped surrogate objective (Eq. 7) is standard; no convergence analysis specific to this setting is given.
  • domain assumption The 288-object training prompt distribution, though filtered to avoid direct overlap, is representative of the GenEval evaluation distribution.
    If the filter is leaky, reported GenEval gains could reflect train/test similarity rather than true compositional generalization (Section 4.1, Training Data).

how reviews work

0 comments
Cite this review

Pith. "Pith review of RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning." pith.science (2026). https://pith.science/paper/Z334Q7I3

@misc{pith2026250517540,
  author       = {Pith},
  title        = {Pith review of: RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z334Q7I3}},
  note         = {Machine review of arXiv:2505.17540}
}
read the original abstract

Despite recent progress in text-to-image (T2I) generation, existing models often struggle to faithfully capture user intentions from short and under-specified prompts. While prior work has attempted to enhance prompts using large language models (LLMs), these methods frequently generate stylistic or unrealistic content due to insufficient grounding in visual semantics and real-world composition. Inspired by recent advances in reasoning for language model, we propose RePrompt, a novel reprompting framework that introduces explicit reasoning into the prompt enhancement process via reinforcement learning. Instead of relying on handcrafted rules or stylistic rewrites, our method trains a language model to generate structured, self-reflective prompts by optimizing for image-level outcomes. The tailored reward models assesse the generated images in terms of human preference, semantic alignment, and visual composition, providing indirect supervision to refine prompt generation. Our approach enables end-to-end training without human-annotated data. Experiments on GenEval and T2I-Compbench show that RePrompt significantly boosts spatial layout fidelity and compositional generalization across diverse T2I backbones, establishing new state-of-the-art results.

Figures

Figures reproduced from arXiv: 2505.17540 by the authors.

Figure 1
Figure 1. Given the user prompt "a photo of a couch below a vase", existing models like DELL￾E3 generate rich language descriptions but often produce unrealistic or physically implausible compositions. In contrast, our RePrompt performs explicit chain-of-thought reasoning to resolve spatial relations, resulting in enhanced prompts that guide text-to-image models towards realistic and semantically aligned generations. linguist… view at source ↗
Figure 2
Figure 2. Overview of the proposed RePrompt. For each input prompt, RePrompt generates multiple [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The Visual-Reasoning Reward. A central component of our framework is the T2I RePrompt Reward Model—an ensemble, image￾grounded reward function specifically designed for the prompt refinement task in T2I generation. In contrast to generic reward functions used in natural language or vision tasks, our reward model is co￾developed with the objective of enhancing reasoning￾driven prompt construction. It explicitly evalu… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Impact of our method across different base T2I models on the GenEval benchmark. Our [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on compositional prompts. Compared to vanilla T2I models, our [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Training curve of our RePrompt during reinforcement learning. The curve shows steady [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Training curve with different visual reasoning rewards. Using a single reward model [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on compositional prompts (Position). We compare DELL-E3, the [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results on compositional prompts (Two-object). We show comparisons among [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results on compositional prompts (Color). We present qualitative comparisons [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results on compositional prompts (Attribute binding). We show examples where [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 10 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2(3):8, 2023

    2023

  3. [3]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023

    arXiv 2023

  4. [4]

    BeautifulPrompt: Towards Automatic Prompt Engineering for Text-to-Image Synthesis

    Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beauti- fulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023

    work page Pith review arXiv 2023

  5. [5]

    Training-free regional prompting for diffusion transformers

    Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompting for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024

    arXiv 2024

  6. [6]

    Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024

    2024

  7. [7]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025

    arXiv 2025

  8. [8]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

    2024

Show all 63 references
  1. [9]

    Video-r1: Reinforcing video reasoning in mllms

    Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms. arXiv preprint arXiv:2503.21776, 2025

    2025 arXiv

  2. [10]

    Geneval: An object-focused framework for evaluating text-to-image alignment

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132–52152, 2023

    2023

  3. [11]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  4. [12]

    Versat2i: Improving text-to-image models with versatile reward

    Jianshu Guo, Wenhao Chai, Jie Deng, Hsiang-Wei Huang, Tian Ye, Yichen Xu, Jiawei Zhang, Jenq-Neng Hwang, and Gaoang Wang. Versat2i: Improving text-to-image models with versatile reward. arXiv preprint arXiv:2403.18493, 2024

    2024 arXiv

  5. [13]

    Can we generate images with cot? let’s verify and reinforce image generation step by step

    Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. Can we generate images with cot? let’s verify and reinforce image generation step by step. arXiv preprint arXiv:2501.13926, 2025

    2025 arXiv

  6. [14]

    A simple and effective reinforcement learning method for text-to-image diffusion fine-tuning

    Shashank Gupta, Chaitanya Ahuja, Tsung-Yu Lin, Sreya Dutta Roy, Harrie Oosterhuis, Maarten de Rijke, and Satya Narayan Shukla. A simple and effective reinforcement learning method for text-to-image diffusion fine-tuning. arXiv preprint arXiv:2503.00897, 2025

    2025

  7. [15]

    Optimizing prompts for text-to-image generation

    Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36:66923–66939, 2023

    2023

  8. [16]

    T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723–78747, 2023. 10

    2023

  9. [17]

    Boosting mllm reasoning with text-debiased hint-grpo.arXiv preprint arXiv:2503.23905, 2025

    Qihan Huang, Long Chan, Jinlong Liu, Wanggui He, Hao Jiang, Mingli Song, Jingyuan Chen, Chang Yao, and Jie Song. Boosting mllm reasoning with text-debiased hint-grpo.arXiv preprint arXiv:2503.23905, 2025

    2025 arXiv

  10. [18]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  11. [19]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  12. [20]

    T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot

    Dongzhi Jiang, Ziyu Guo, Renrui Zhang, Zhuofan Zong, Hao Li, Le Zhuo, Shilin Yan, Pheng- Ann Heng, and Hongsheng Li. T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot. arXiv preprint arXiv:2505.00703, 2025

    2025 arXiv

  13. [21]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:36652–36663, 2023

    2023

  14. [22]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux , 2024

    2024

  15. [23]

    Aligning text-to-image models using human feedback

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023

    2023 arXiv

  16. [24]

    Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation

    Seung Hyun Lee, Yinxiao Li, Junjie Ke, Innfarn Yoo, Han Zhang, Jiahui Yu, Qifei Wang, Fei Deng, Glenn Entis, Junfeng He, et al. Parrot: Pareto-optimal multi-reward reinforcement learning framework for text-to-image generation. In European Conference on Computer Vision, pages 4...

    2024

  17. [25]

    Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning

    Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning. arXiv preprint arXiv:2504.06958, 2025

    2025 arXiv

  18. [26]

    Rich human feedback for text-to-image generation

    Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  19. [27]

    Playground v3: Im- proving text-to-image alignment with deep-fusion large language models

    Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Chase Lambert, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Im- proving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024

    2024 arXiv

  20. [28]

    Ui-r1: Enhancing action prediction of gui agents by reinforcement learning

    Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Guanjing Xiong, and Hongsheng Li. Ui-r1: Enhancing action prediction of gui agents by reinforcement learning. arXiv preprint arXiv:2503.21620, 2025

    2025 arXiv

  21. [29]

    Inference-time scaling for diffusion models beyond scaling denoising steps

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732, 2025

    2025 arXiv

  22. [30]

    Deepperception: Advancing r1-like cognitive visual perception in mllms for knowledge-intensive visual grounding

    Xinyu Ma, Ziyang Ding, Zhicong Luo, Chi Chen, Zonghao Guo, Derek F Wong, Xiaoyi Feng, and Maosong Sun. Deepperception: Advancing r1-like cognitive visual perception in mllms for knowledge-intensive visual grounding. arXiv preprint arXiv:2503.12797, 2025

    2025

  23. [31]

    Improving text-to-image consistency via automatic prompt optimization

    Oscar Mañas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdzal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024

    2024 arXiv

  24. [32]

    Dynamic prompt optimizing for text-to-image generation

    Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26627–26636, 2024. 11

    2024

  25. [33]

    Personalized and sequential text-to-image generation

    Ofir Nabati, Guy Tennenholtz, ChihWei Hsu, Moonkyung Ryu, Deepak Ramachandran, Yinlam Chow, Xiang Li, and Craig Boutilier. Personalized and sequential text-to-image generation. arXiv preprint arXiv:2412.10419, 2024

    2024 arXiv

  26. [34]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  27. [35]

    Diffusiongpt: Llm-driven text-to-image generation system

    Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024

    2024

  28. [36]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  29. [37]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  30. [38]

    Faircot: Enhancing fairness in diffu- sion models via chain of thought reasoning of multimodal language models

    Zahraa Al Sahili, Ioannis Patras, and Matthew Purver. Faircot: Enhancing fairness in diffu- sion models via chain of thought reasoning of multimodal language models. arXiv preprint arXiv:2406.09070, 2024

    2024 arXiv

  31. [39]

    Vlm-r1: A stable and generalizable r1-style large vision-language model

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615, 2025

    2025 arXiv

  32. [40]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025

    2025 arXiv

  33. [41]

    Reft: Reason- ing with reinforced fine-tuning

    Luong Trung, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reason- ing with reinforced fine-tuning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7601–7614, 2024

    2024

  34. [42]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and...

    2024

  35. [43]

    Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and rl

    Junke Wang, Zhi Tian, Xun Wang, Xinyu Zhang, Weilin Huang, Zuxuan Wu, and Yu-Gang Jiang. Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and rl. arXiv preprint arXiv:2504.11455, 2025

    2025 arXiv

  36. [44]

    Mint: Multi-modal chain of thought in unified generative models for enhanced image generation

    Yi Wang, Mushui Liu, Wanggui He, Longxiang Zhang, Ziwei Huang, Guanghao Zhang, Fangxun Shu, Zhong Tao, Dong She, Zhelun Yu, et al. Mint: Multi-modal chain of thought in unified generative models for enhanced image generation. arXiv preprint arXiv:2503.01298, 2025

    2025

  37. [45]

    Genartist: Multimodal llm as an agent for unified image generation and editing

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. Advances in Neural Information Processing Systems, 37:128374–128395, 2024

    2024

  38. [46]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  39. [47]

    Self-correcting llm-controlled diffusion models

    Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327–6336, 2024. 12

    2024

  40. [48]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng Yu, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, et al. Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv preprint arXiv:2501.18427, 2025

    2025 arXiv

  41. [49]

    Show-o: One single trans- former to unify multimodal understanding and generation

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single trans- former to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024

    2024 arXiv

  42. [50]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023

    2023

  43. [51]

    Dancegrpo: Unleashing grpo on visual generation

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818, 2025

    2025 arXiv

  44. [52]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  45. [53]

    Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms

    Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In Forty-first International Conference on Machine Learning, 2024

    2024

  46. [54]

    A dense reward view on aligning text-to- image diffusion with preference

    Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to- image diffusion with preference. arXiv preprint arXiv:2402.08265, 2024

    2024 arXiv

  47. [55]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization. arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  48. [56]

    Idea2img: Iterative self-refinement with gpt-4v for automatic image design and generation

    Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Idea2img: Iterative self-refinement with gpt-4v for automatic image design and generation. In European Conference on Computer Vision, pages 167–184. Springer, 2024

    2024

  49. [57]

    Tipo: Text to image with text presampling for prompt optimization

    Shih-Ying Yeh, Sang-Hyun Park, Giyeong Oh, Min Song, and Youngjae Yu. Tipo: Text to image with text presampling for prompt optimization. arXiv preprint arXiv:2411.08127, 2024

    2024

  50. [58]

    Perception-r1: Pioneering perception policy with reinforcement learning

    En Yu, Kangheng Lin, Liang Zhao, Jisheng Yin, Yana Wei, Yuang Peng, Haoran Wei, Jian- jian Sun, Chunrui Han, Zheng Ge, et al. Perception-r1: Pioneering perception policy with reinforcement learning. arXiv preprint arXiv:2504.07954, 2025

    2025 arXiv

  51. [59]

    Learning to sample effective and diverse prompts for text-to-image generation

    Taeyoung Yun, Dinghuai Zhang, Jinkyoo Park, and Ling Pan. Learning to sample effective and diverse prompts for text-to-image generation. arXiv preprint arXiv:2502.11477, 2025

    2025 arXiv

  52. [60]

    R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization. arXiv preprint arXiv:2503.12937, 2025

    2025 arXiv

  53. [61]

    Itercomp: Iterative composition-aware feedback learning from model gallery for text-to-image generation

    Xinchen Zhang, Ling Yang, Guohao Li, Yaqi Cai, Jiake Xie, Yong Tang, Yujiu Yang, Mengdi Wang, and Bin Cui. Itercomp: Iterative composition-aware feedback learning from model gallery for text-to-image generation. arXiv preprint arXiv:2410.07171, 2024

    arXiv

  54. [62]

    Layercraft: Enhancing text-to-image generation with cot reasoning and layered object integration

    Yuyao Zhang, Jinghao Li, and Yu-Wing Tai. Layercraft: Enhancing text-to-image generation with cot reasoning and layered object integration. arXiv preprint arXiv:2504.00010, 2025

    2025

  55. [63]

    Colors" (from 0.83 to 0.87) and

    Hanyang Zhao, Haoxian Chen, Ji Zhang, David D Yao, and Wenpin Tang. Scores as actions: a framework of fine-tuning diffusion models by continuous-time reinforcement learning. arXiv preprint arXiv:2409.08400, 2024. 13 A Variance Reduction via Structured Reasoning In this appendi...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.