Pith. sign in

REVIEW 5 major objections 6 minor 9 cited by

Draw ALL Your Imagine: A Holistic Benchmark and Agent Framework for Complex Instruction-based Image Generation

T0 review · 5 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 training-free agent that decomposes complex image prompts into layered scenes out-scores GPT-4o on a new 500-prompt benchmark.

desk verdict Useful benchmark, unproven SOTA claim: the Plan2Gen–GPT-4o gap is 0.02–0.03 on an unvalidated MLLM scale, and Table 3 has a copy-paste error. read the letter →

arxiv 2505.24787 v1 pith:G43GURTG submitted 2025-05-30 cs.CV cs.CL

Yucheng Zhou , Jiahao Yuan , Qianning Wang This is my paper
classification cs.CVcs.CL
keywords LongBench-T2IPlan2Gentext-to-imagegenerationcomplexinstructionfollowingbenchmarkagentframeworkmultimodalevaluationscenedecomposition
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

The paper argues that existing text-to-image benchmarks miss what happens when a prompt spans many objects, attributes, spatial relations, and visual effects, and that this gap hides a fixable weakness in current models. It introduces LongBench-T2I, a set of 500 long, human-filtered prompts annotated across nine visual dimensions, and uses it to measure how well models obey complex instructions. On top of that, it proposes Plan2Gen, a training-free agent that asks a large language model to split a prompt into background, midground, and foreground layers, generates each layer in sequence, and validates and refines each layer before moving on. The paper's central empirical claim is that Plan2Gen, built on Gemini-2.0, reaches the highest average scores on this benchmark (3.73 with Gemini-2.0-Flash as judge, 3.41 with InternVL3-78B as judge), edging out GPT-4o. The significance would be that complex instruction following can be improved through planning and self-correction alone, without retraining any generator.

What carries the argument

The load-bearing object is the Plan2Gen agent pipeline rather than any new generative model. Its core mechanism is progressive layered image generation: a language model reads the full instruction and separates it into background, midground, and foreground sub-prompts; each layer is generated conditioned on the approved previous layers; a validator LLM checks the layer against its sub-prompt; and mismatches trigger prompt-level refinements, with a maximum of three refinement steps per layer. The benchmark side rests on nine extracted visual elements (objects, background, color, texture, lighting, text/symbol, composition, pose, special effects) that instructions are parsed into and that evaluator MLLMs score.

What would settle it

A human-rating study on a random subset of the 500 benchmark images: if human judges disagree with Gemini-2.0-Flash or InternVL3-78B on the nine dimensions, or if human aggregate rankings place GPT-4o above Plan2Gen, the central claim of superiority is falsified. A simpler check: re-run the same Plan2Gen pipeline with a different open-source planner and see whether the average-score gap to GPT-4o persists.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a planning-driven agent is currently the best way to turn long, multi-faceted text instructions into images. Plan2Gen decomposes each 500-token-scale prompt into three compositional layers, generates them progressively with each layer conditioned on the previous ones, and runs an LLM-based validation-refinement loop per layer. In the paper's measurements this pipeline outperforms every standalone model tested, including GPT-4o, while the planning step alone explains much of the gain: swapping the planner from Gemini-2.0-Flash to Gemini-1.5-Flash-8b drops the average from 3.73 to 3.42, and plan consistency scores correlate with final image quality. The paper also reports a human evaluation in which Plan2Gen wins more often than it loses against GPT-4o in eight of nine dimensions, losing only on object fidelity.

Load-bearing premise

The rankings, including the claim that Plan2Gen beats GPT-4o, rest on the assumption that two multimodal LLM judges score the nine visual dimensions on a 1-5 scale as reliably as a human would; the paper does not calibrate those automated scores against human ratings of the same images.

Editorial extensions

If this is right

  • If Plan2Gen's results hold, any existing text-to-image model can gain on complex instructions purely by adding an LLM planning and refinement loop, with no weight updates.
  • The nine-dimension scoring scheme gives model developers a diagnostic: a model's average can hide very different per-dimension failures, e.g., 'Text' is hardest for nearly every model.
  • The benchmark's 500 long prompts provide a standard testbed for long-context instruction following, complementing short-prompt benchmarks such as DrawBench and T2I-CompBench.
  • The refinement analysis suggests a sweet spot: about three validation-refinement cycles per layer; more iterations risk over-correction.
  • The perplexity-vs-score analysis implies that language understanding alone does not guarantee image fidelity, especially in smaller models.

Reading between the lines

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

  • One implication the paper leaves implicit is that the automated MLLM scores should be validated against human ratings on the same images; until that calibration exists, the exact ranking margin over GPT-4o is provisional.
  • The layered decomposition assumes a scene decomposes cleanly into background, midground, and foreground; prompts where elements interpenetrate or where ordering is ambiguous may stress the agent, suggesting a test set variant with deliberately entangled spatial relations.
  • The same plan-and-refine loop could transfer to other generative domains such as video or 3D scenes, where a storyboard-style decomposition plays the role of the three layers.
  • The counter-intuitive positive perplexity-score correlation for Janus-Pro-1B points to a testable extension: measuring whether instruction-tuning on low-perplexity complex prompts closes the translation gap.
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

5 major / 6 minor

Summary. LongBench-T2I is introduced as a benchmark of 500 complex text-to-image prompts annotated along nine visual dimensions, together with an automated evaluation toolkit that uses two MLLMs (Gemini-2.0-Flash and InternVL3-78B) to score generated images per dimension. The paper also presents Plan2Gen, a training-free agent framework that decomposes a long instruction into background, midground, and foreground sub-prompts and then generates, validates, and refines each layer in sequence. Experiments compare a range of diffusion and autoregressive T2I models on the benchmark, reporting that Plan2Gen achieves the highest average scores, slightly above GPT-4o, and a human pairwise preference study on a subset of prompts.

Significance. If the benchmark is valid and the automated scores are calibrated, LongBench-T2I would fill a real gap in evaluating complex instruction-following, and the public release of data, code, and an open-source evaluator is a practical strength that should enable future comparisons. Plan2Gen is a plausible training-free baseline. However, the central comparative claim is currently supported only by unvalidated automated scores with tiny absolute differences and no uncertainty quantification, so the contribution's significance is conditional on a substantially strengthened evaluation.

major comments (5)
  1. [Section 5.2, Tables 2 and 3] The headline result that Plan2Gen outperforms GPT-4o rests on average score differences of 0.03 (Gemini-2.0-Flash) and 0.02 (InternVL3-78B) on a 1-5 scale. No variance, confidence interval, or significance test is reported for any score, so it is impossible to tell whether this ordering is reliable. Because Plan2Gen uses Gemini-2.0-Flash both as generator/refiner and as evaluator in Table 2, there is also a same-family evaluator risk. Please report per-prompt paired differences with a paired significance test (e.g., Wilcoxon signed-rank) and bootstrap confidence intervals, and ideally a third-party evaluator.
  2. [Sections 5.1 and 5.3] The nine-dimension 1-5 scores produced by Gemini-2.0-Flash and InternVL3-78B are never validated against human ratings on the same rubric. The human evaluation in Section 5.3 is a pairwise preference between Plan2Gen and two baselines, which does not calibrate the absolute dimension scores used in Tables 2-3. Without a human-calibration study showing that the automated scores agree with human judgments (e.g., Spearman correlation or ICC per dimension), the magnitudes and rankings in Tables 2-3 are ungrounded. Please add such a study on a representative sample.
  3. [Section 5.5, Table 4] The maximum refinement steps for Plan2Gen is set to 3 because that value gives the highest average score on LongBench-T2I, which is the same test set used for the final comparison in Tables 2-3. This is a form of test-set hyperparameter selection and biases the reported Plan2Gen scores upward. Please either tune the number of refinement steps on a separate validation split (e.g., a held-out subset of prompts) or present the full sensitivity curve as the primary result.
  4. [Section 5.2, Table 3 and Figure 5] The row for Infinity-2B-reg is a verbatim duplicate of the Nexus-Gen row (e.g., Obj. 1.78, Backg. 2.88, Color 2.84), which is impossible for two distinct models; this indicates a data handling error and undermines confidence in the reported numbers. Similarly, Figure 5's right panel appears to contain corrupted text (the '/uni0000...' sequences), making the plot unreadable. Please correct the table and replace the figure with a properly rendered version.
  5. [Sections 4 and 5.2] Plan2Gen is not matched to baselines for inference compute. It performs up to three generation-validation-refinement cycles per layer across three layers, while baseline models such as GPT-4o are evaluated as single-pass generators. The observed advantage could therefore be due to additional inference compute rather than the planning mechanism. Please include a compute-matched baseline (e.g., GPT-4o with self-correction or repeated sampling) and report the average number of generation calls per prompt for each method.
minor comments (6)
  1. [Section 3.4] The human review is conducted by only two graduate students; please report inter-annotator agreement (e.g., Cohen's kappa) to demonstrate reliability of the filtering process.
  2. [Section 5.2, point (5)] The claimed 'evaluator consistency' is not quantified; please report a rank correlation (e.g., Spearman) between the two evaluators' model rankings.
  3. [Section 5.4] The correlation between perplexity and evaluation score is described qualitatively; please report the correlation coefficient, p-value, and number of samples, and clarify whether the fit is on per-prompt or per-model data.
  4. [General] The paper uses an ACM template with placeholder conference information (e.g., 'Conference acronym XX, June 03-05, 2018' and '7 pages'); please update to the correct venue details.
  5. [Figure 1] Figure 1 is visually dense; consider separating the prompt example from the evaluation framework diagram to improve readability.
  6. [Abstract and Conclusion] The abstract and conclusion state that Plan2Gen 'outperforms' other models; given the statistical caveats above, please qualify these statements (e.g., 'achieves the highest average score in our evaluation').

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the benchmark, agent framework, and evaluation are empirically connected rather than defined into equivalence.

full rationale

The paper's derivation chain is empirical rather than formal: LongBench-T2I is constructed via LLM generation, element extraction, automated review, and human review; Plan2Gen is a separate training-free agent that decomposes prompts into background/midground/foreground layers; and evaluation is performed by MLLMs scoring nine visual dimensions. No equation or construction forces Plan2Gen's scores to equal the benchmark's design choices. The nine evaluation dimensions are used as the scoring rubric, but Plan2Gen's decomposition into three spatial layers is a different structure, so the benchmark does not define the method's output by construction. The headline superiority over GPT-4o rests on small MLLM score differences, and there is a possible same-family evaluator risk since Gemini-2.0-Flash is both the Plan2Gen planner and one of the judges; these are validity and correctness concerns, not circularity under the specified patterns. The self-citations [32]–[34] appear only in related-work context and do not carry any load-bearing premise, so they do not raise the circularity score. An internal data error in Table 3 (Infinity-2B-reg row identical to Nexus-Gen) further undermines confidence in specific numbers, but again this is not a circular reduction. Overall, the central claims are independently grounded by external baselines and a human pairwise study, albeit with limitations, so the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper is empirical and does not introduce physical entities or fitted equations. Design choices such as the nine dimensions and the 500-prompt size are taxonomies, not fitted parameters. The only clear tuned numeric is the maximum refinement steps, selected on the evaluation test set.

free parameters (1)
  • max_refine_steps = 3
    Selected by sweeping 1, 3, 5, and 7 on the LongBench-T2I test set (Table 4) and choosing the best average score; this is a test-set-tuned hyperparameter.
assumptions (3)
  • domain assumption MLLM evaluators (Gemini-2.0-Flash, InternVL3-78B) provide valid 1-5 scores of instruction following in nine dimensions.
    The entire comparative evaluation relies on this; the paper does not validate the automated scores against human ratings (Section 5.1, Tables 2-3).
  • domain assumption LLM-generated prompts filtered by two graduate students are representative of complex instructions.
    The benchmark is built via an LLM pipeline with human review; no inter-annotator agreement or external corpus comparison is provided (Section 3).
  • ad hoc to paper The nine visual dimensions jointly cover complex instruction-following.
    The dimensions are defined by the authors and used both to construct prompts and to evaluate them, creating a circular dependency (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Draw ALL Your Imagine: A Holistic Benchmark and Agent Framework for Complex Instruction-based Image Generation." pith.science (2026). https://pith.science/paper/G43GURTG

@misc{pith2026250524787,
  author       = {Pith},
  title        = {Pith review of: Draw ALL Your Imagine: A Holistic Benchmark and Agent Framework for Complex Instruction-based Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G43GURTG}},
  note         = {Machine review of arXiv:2505.24787}
}
read the original abstract

Recent advancements in text-to-image (T2I) generation have enabled models to produce high-quality images from textual descriptions. However, these models often struggle with complex instructions involving multiple objects, attributes, and spatial relationships. Existing benchmarks for evaluating T2I models primarily focus on general text-image alignment and fail to capture the nuanced requirements of complex, multi-faceted prompts. Given this gap, we introduce LongBench-T2I, a comprehensive benchmark specifically designed to evaluate T2I models under complex instructions. LongBench-T2I consists of 500 intricately designed prompts spanning nine diverse visual evaluation dimensions, enabling a thorough assessment of a model's ability to follow complex instructions. Beyond benchmarking, we propose an agent framework (Plan2Gen) that facilitates complex instruction-driven image generation without requiring additional model training. This framework integrates seamlessly with existing T2I models, using large language models to interpret and decompose complex prompts, thereby guiding the generation process more effectively. As existing evaluation metrics, such as CLIPScore, fail to adequately capture the nuances of complex instructions, we introduce an evaluation toolkit that automates the quality assessment of generated images using a set of multi-dimensional metrics. The data and code are released at https://github.com/yczhou001/LongBench-T2I.

Figures

Figures reproduced from arXiv: 2505.24787 by the authors.

Figure 1
Figure 1. LongBench-T2I benchmark evaluates text-to-image models on their ability to follow complex instructions. The framework extracts nine key elements from each instruction to systematically assess the detail of the generated image. Abstract Recent advancements in text-to-image (T2I) generation have en￾abled models to produce high-quality images from textual descrip￾tions. However, these models often struggle with complex… view at source ↗
Figure 2
Figure 2. The multi-stage pipeline for generating the LongBench-T2I benchmark. skills per instance. DPG-Bench [10] expands compositional den￾sity through template-based multi-object prompts, yet its auto￾generated language lacks the descriptive nuance and narrative structure required for evaluating long-form instruction following. However, these are limited by brevity, narrow diversity, or synthetic prompts. In contrast, as s… view at source ↗
Figure 3
Figure 3. The Plan2Gen Agent generates images from complex long instructions by first decomposing the scene into background, midground, and foreground prompts, then progressively generating and refining each layer with validation to ensure alignment with the original description. Visual Element Richness: The diversity, detail, and specificity of the described visual components. ❷ Structural and Compositional Complexity: The i… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparative performance evaluation across nine visual dimensions when our model is compared against Gemini-2.0 (left) and GPT-4o (right). Each horizontal bar shows the proportion of images where our model was rated as better (Win), equivalent (Tie), or worse (Lose). 10…
Figure 5
Figure 5. Figure 5: (Left) Instruction perplexity vs. average evaluation score for Janus-Pro models. Each point represents a sample. The red line indicates the linear fit in log10 perplexity space. (Right) Relationship Between Plan Consistency and Generation Quality. higher win rates in a…

Discussion (0). Sign in to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Spectral Origins of the Self-Correction Blind Spot in Autoregressive Generation

    cs.LG 2026-07 conditional novelty 5.5 of 10

    Self-correction blind spots in residual-stream autoregressive models arise iff the product of attention Jacobians has spectral radius ≥1, with a sharp marker threshold and RL coupling condition derived from that radius.

  2. DocRefine: An Intelligent Framework for Scientific Document Understanding and Content Optimization based on Multimodal Large Model Agents

    cs.CV 2025-08 reject novelty 4.0 of 10

    A multi-agent GPT-4o framework for editing scientific PDFs reports higher semantic consistency, layout fidelity, and instruction adherence than three baselines on DocEditBench.

  3. LumiGen: An LVLM-Enhanced Iterative Framework for Fine-Grained Text-to-Image Generation

    cs.LG 2025-08 reject novelty 4.0 of 10

    An LVLM-driven iterative text-to-image framework whose claimed performance scores are explicitly labeled fictitious, so no empirical result is established.

  4. Harnessing RLHF for Robust Unanswerability Recognition and Trustworthy Response Generation in LLMs

    cs.CL 2025-07 reject novelty 4.0 of 10

    SALU, a multi-task fine-tuning and confidence-guided RLHF method, reduces hallucinated answers on unanswerable Chinese CIR questions to 1.3 percent on the authors' private dataset.

  5. CIMR: Contextualized Iterative Multimodal Reasoning for Robust Instruction Following in LVLMs

    cs.LG 2025-07 reject novelty 4.0 of 10

    CIMR, an iterative reasoning wrapper around LLaVA-1.5-7B, reports 91.5% task completion on a newly constructed but unreleased synthetic MAP dataset, above GPT-4V at 89.2%.

  6. LVLM-Composer's Explicit Planning for Image Generation

    cs.CV 2025-07 reject novelty 4.0 of 10

    An image generation model that explicitly plans objects, attributes, locations, and relations before synthesizing the image, with reported gains on LongBench-T2I that cannot be verified from the paper.

  7. MM-FusionNet: Context-Aware Dynamic Fusion for Multi-modal Fake News Detection with Large Vision-Language Models

    cs.CR 2025-08 reject novelty 3.0 of 10

    MM-FusionNet uses bi-directional cross-modal attention and a dynamic gating network to weight text and image features for fake news detection, reporting 0.938 F1 on the private LMFND dataset.

  8. Unlocking Compositional Control: Self-Supervision for LVLM-Based Image Generation

    cs.CV 2025-07 reject novelty 3.0 of 10

    Hi-SSLVLM combines hierarchical self-captioning, internal sub-prompt planning, and a CLIP-based consistency loss, and reports judged compositional fidelity gains of roughly 0.04 to 0.09 points that no significance tes...

  9. Large Language Models for Zero-Shot Multicultural Name Recognition

    cs.CL 2025-07 reject novelty 3.0 of 10

    A prompt-tuned LLM with data augmentation and cultural context prompts reportedly recognizes multicultural names at 93.1% accuracy and unseen names at 89.5%, but the evidence is not reproducible.

Reference graph

Works this paper leans on

36 extracted references · 7 canonical work pages · cited by 9 Pith papers

  1. [1]

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

    2023

  2. [2]

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

    arXiv 2025

  3. [3]

    Google DeepMind. 2025. Gemini 2.0 Flash Experimental API (Image Generation). https://aistudio.google.com/prompts/new_chat Experimental model accessed via Gemini API, May 2025

    work page 2025

  4. [4]

    Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. 2023. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023)

    arXiv 2023

  5. [5]

    Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Shilin Yan, Hao Tian, Xingyu Zeng, Rui Zhao, Jifeng Dai, et al . 2025. Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing. arXiv preprint arXiv:2503.10639 (2025)

    arXiv 2025

  6. [6]

    Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Peng Gao, Hongsheng Li, and Pheng-Ann Heng. 2025. Can We Generate Images with CoT? Let’s Verify and Reinforce Image Generation Step by Step. arXiv preprint arXiv:2501.13926 (2025)

    arXiv 2025

  7. [7]

    Meera Hahn, Wenjun Zeng, Nithish Kannen, Rich Galt, Kartikeya Badola, Been Kim, and Zi Wang. 2024. Proactive Agents for Multi-Turn Text-to-Image Genera- tion Under Uncertainty. arXiv preprint arXiv:2412.06771 (2024)

    arXiv 2024

  8. [8]

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. 2024. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431 (2024)

    arXiv 2024

Show all 36 references
  1. [9]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. InEMNLP

    2021

  2. [10]

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. 2024. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135 (2024)

    2024 arXiv

  3. [11]

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

    2023

  4. [12]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    2024 arXiv

  5. [13]

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

    2025 arXiv

  6. [14]

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

    2024

  7. [15]

    Xinyao Liao, Wei Wei, Xiaoye Qu, and Yu Cheng. 2025. Step-level Reward for Free in RL-based T2I Diffusion Model Fine-tuning. arXiv preprint arXiv:2505.19196 (2025)

    2025 arXiv

  8. [16]

    Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. 2024. Lumina-mgpt: Illuminate flexible photorealistic text- to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657 (2024)

    2024 arXiv

  9. [17]

    Mayu Otani, Riku Togashi, Yu Sawai, Ryosuke Ishigami, Yuta Nakashima, Esa Rahtu, Janne Heikkilä, and Shin’ichi Satoh. 2023. Toward verifiable and repro- ducible human evaluation for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2023

  10. [18]

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

    2024

  11. [19]

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

    2022

  12. [20]

    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. 2022. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural infor...

    2022

  13. [21]

    Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. 2019. Objects365: A large-scale, high-quality dataset for ob- ject detection. In Proceedings of the IEEE/CVF international conference on computer vision. 8430–8439

    2019

  14. [22]

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. 2024. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525 (2024)

    2024 arXiv

  15. [23]

    Maria Mihaela Trusca, Wolf Nuyts, Jonathan Thomm, Robert Honig, Thomas Hofmann, Tinne Tuytelaars, and Marie-Francine Moens. 2024. Object-attribute binding in text-to-image generation: Evaluation and control. arXiv preprint arXiv:2404.13766 (2024)

    2024 arXiv

  16. [24]

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. 2024. Diffusion model alignment using direct preference optimization. InProceedings of the IEEE/CVF Conference on Computer Visio...

    2024

  17. [25]

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

    2025

  18. [26]

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

    2024

  19. [27]

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

    2024

  20. [28]

    Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. 2024. Deep reward supervisions for tuning text-to-image diffusion models. In European Conference on Computer Vision. Springer, 108–124

    2024

  21. [29]

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. 2024. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340 (2024)

    2024 arXiv

  22. [30]

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

    2023

  23. [31]

    Hong Zhang, Zhongjie Duan, Xingjun Wang, Yingda Chen, Yuze Zhao, and Yu Zhang. 2025. Nexus-Gen: A Unified Model for Image Understanding, Generation, and Editing. arXiv preprint arXiv:2504.21356 (2025)

    2025 arXiv

  24. [32]

    Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. 2024. Visual in- context learning for large vision-language models.arXiv preprint arXiv:2402.11574 (2024)

    2024 arXiv

  25. [33]

    Yucheng Zhou, Jianbing Shen, and Yu Cheng. 2025. Weak to strong generalization for large language models with multi-capabilities. In The Thirteenth International Conference on Learning Representations

    2025

  26. [34]

    Yucheng Zhou, Jihai Zhang, Guanjie Chen, Jianbing Shen, and Yu Cheng. 2024. Less is more: Vision representation compression for efficient video generation with large language models

    2024

  27. [35]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. 2025. InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models. arXiv preprint arXiv:2504.10479 (2025)

    2025 arXiv

  28. [36]

    Xianwei Zhuang, Yuxin Xie, Yufan Deng, Liming Liang, Jinghan Ru, Yuguo Yin, and Yuexian Zou. 2025. VARGPT: Unified Understanding and Generation in a Visual Autoregressive Multimodal Large Language Model. arXiv preprint arXiv:2501.12327 (2025). 7

    2025 arXiv

Pith tools

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