Pith. sign in

REVIEW 3 major objections 6 minor 5 cited by

OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation

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 OneIG-Bench's central claim is that text-to-image quality should be measured as six separately scoreable dimensions, and the benchmark's automated metrics make that decomposition reproducible.

desk verdict Solid multi-dimensional T2I benchmark with a valuable bilingual dataset, but the reasoning metric is self-referential and the human-validation claim is unsupported. read the letter →

arxiv 2506.07977 v3 pith:STWCDIS6 submitted 2025-06-09 cs.CV

Jingjing Chang , Yixiao Fang , Peng Xing , Shuhan Wu , Wei Cheng , Rui Wang , Xianfang Zeng , Gang Yu
show 1 more author
Hai-Bao Chen
This is my paper
classification cs.CV
keywords text-to-imagegenerationbenchmarkevaluationsemanticalignmenttextrenderingknowledgeandreasoningstylizationmultilingualismdiversitymetric
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

This paper introduces OneIG-Bench, a text-to-image benchmark built from 2,440 prompts, including a 1,320-prompt Chinese subset, organized into six categories: General Object, Portrait, Anime and Stylization, Text Rendering, Knowledge and Reasoning, and Multilingualism. Its central claim is that model quality is not one scalar: the same model can rank first on reasoning but poorly on text rendering, so evaluation should be decomposed into dimensions, each with its own automated metric. The paper reports scores for a range of open-source and closed-source models and argues that the framework separates strengths and bottlenecks, for example isolating GPT-4o's and Imagen4's lead in reasoning and alignment from Seedream 3.0's lead in exact text rendering. It also enables flexible evaluation, letting users generate images only for selected subsets rather than the full prompt set, and it releases the code and data. If the benchmark is right, it gives researchers a reproducible, dimension-wise ranking that earlier single-purpose benchmarks could not provide.

What carries the argument

The machinery is the six-category taxonomy plus a purpose-built metric for each category. The named objects are: the question dependency graph from DSG for semantic alignment; the composite text score $S_{\text{text}}=1-\min(\phi,\mathrm{ED})(1-\mathrm{CR})(1-\mathrm{WAC})/\phi$, with $\phi=100$ for English and $\phi=50$ for Chinese; GPT-4o reference answers paired with LLM2CLIP cosine similarity for knowledge and reasoning; a dual style-encoding pipeline using CSD and a CLIP-based OneIG encoder, taking the maximum cosine similarity to reference images; and DreamSim-based pairwise diversity computed over multiple images per prompt. The framework's flexibility comes from treating each dimension as an independent subset, so a user can run one evaluation slice without generating images for the entire prompt set.

What would settle it

Take the 225 Knowledge and Reasoning prompts, replace GPT-4o's reference answers with answers written by a different large language model or by human experts, recompute the reasoning scores, and check whether the ranking of models changes; if leading models swap positions, the reasoning dimension is not measuring a stable capability.

Watch

Extended reading notes

Core claim

The central discovery is a measurement decomposition: OneIG-Bench claims to make fine-grained evaluation of text-to-image models reproducible across the capabilities that matter in practice. It provides curated prompts, roughly 200 per category, drawn from real-world inputs, deduplicated, and rewritten into short, middle, and long lengths, with each category assigned a metric designed for that capability. Alignment is scored by decomposing a prompt into a dependency graph and verifying root and leaf questions against the generated image with a vision-language model. Text rendering is scored by edit distance, exact completion rate, and word accuracy, fused into a single text score. Reasoning is scored by the cosine similarity between the generated image and a GPT-4o-generated textual answer, computed with LLM2CLIP. Style is scored by averaging two style encoders' similarities to reference images, and diversity is scored by pairwise DreamSim distance among images from the same prompt. The results rank closed-source models first on most dimensions and show that no single model dominates all six.

Load-bearing premise

The reasoning score assumes that GPT-4o's written answers are the correct reference for what a reasoning image should contain, and that the text-image similarity score used to compare those answers with generated images really measures reasoning quality, even though GPT-4o is also one of the models being ranked.

Editorial extensions

If this is right

  • Because each dimension is scored independently, a model can be selected by application: exact text rendering, reasoning, style, alignment, or diversity can be compared without conflating qualities into one number.
  • The bilingual comparison exposes a common weakness: most models can barely render Chinese text, while Seedream 3.0 reaches a text score of 0.928 on OneIG-Bench-ZH, so text rendering should be reported separately per script.
  • The edit-distance rule that counts any capitalization mismatch as an error directly lowers visually strong renderers such as GPT-4o, so text rankings should be interpreted with that strictness in mind.
  • The reasoning dimension gives reasoning-oriented text-to-image models a place to show capability that alignment-only benchmarks miss, and the paper's tiered ranking can serve as a target for open-source models.
  • The deliberate 1:2:1 short, middle, and long prompt structure makes it possible to isolate where long-prompt failures begin, since alignment and text scores degrade with prompt length for most models.

Reading between the lines

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

  • The reasoning score may be sensitive to which model writes the reference answer, because GPT-4o supplies the textual ground truth and is also one of the scored image generators; swapping the answer generator for an independent source would test whether the rankings persist.
  • The diversity metric is informative mainly among models with comparable alignment, since outputs that ignore the prompt can look diverse; conditioning diversity on alignment or reporting both jointly would make the comparison more reliable.
  • The same per-dimension construction could be extended to video or multi-image generation by reusing the prompt taxonomy and replacing the image-level metrics with temporal or cross-frame checks.
  • Public release of per-prompt scores, not just averages, would let the community test whether dimension scores are stable under prompt resampling and would expose error patterns that aggregate scores hide.
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 introduces OneIG-Bench, an English-Chinese benchmark containing 2,440 prompts across six categories (General Object, Portrait, Anime & Stylization, Text Rendering, Knowledge & Reasoning, Multilingualism), together with automated metrics for semantic alignment, text rendering, style, diversity, and knowledge/reasoning. The benchmark is validated by evaluating 22 text-to-image models, including open-source, unified multimodal, and proprietary models, and reporting per-dimension scores with fine-grained breakdowns by prompt type and length. The authors make the dataset and code publicly available and highlight the benchmark's flexibility in evaluating user-selected subsets.

Significance. If the proposed metrics are valid, OneIG-Bench fills real gaps in the evaluation landscape: it explicitly targets reasoning, style, text rendering, and multilingual generation, dimensions that earlier benchmarks such as GenEval, DPG-Bench, and T2ICompBench only partially cover. The public release of code and data is a concrete strength that supports reproducibility and cross-model comparison. The most significant caveat is that the Knowledge and Reasoning metric depends on a circular, unvalidated reference, which currently undermines the benchmark's claim to provide a trustworthy multidimensional evaluation. Because the reasoning dimension is a headline contribution, this issue must be resolved before the benchmark can be relied upon.

major comments (3)
  1. [§4.1, Knowledge and Reasoning] The reasoning score uses GPT-4o-generated textual answers as the reference and LLM2CLIP cosine similarity between that text and the generated image as the score. Since GPT-4o image generation is itself one of the evaluated models (Table 2), the evaluation is self-referential: it measures how closely an image matches GPT-4o's own textual expression, not whether the image is factually correct or logically coherent. The consequence is visible in Tables 2 and 3, where GPT-4o receives the highest reasoning scores (0.345 and 0.300). No evidence is provided that LLM2CLIP cosine similarity tracks human judgments of reasoning quality. This is load-bearing because reasoning is one of the six benchmark dimensions and is emphasized in the abstract as a key contribution.
  2. [Limitation section] The single sentence claiming that "our metric rankings align closely with human evaluations" is not accompanied by any protocol, data, or statistical measure. For a benchmark whose purpose is to provide valid automated rankings, the absence of a human validation study is a serious omission. At minimum, the authors should report the number of annotators, the number of prompts sampled, the agreement metric (e.g., Spearman or Kendall correlation), and ideally a breakdown per dimension, with particular attention to the reasoning and style dimensions.
  3. [§4.1, Style] The style score relies on a newly introduced "OneIG style image encoder" fine-tuned from CLIP using CSGO-generated images, but the manuscript provides no details on the training data, architecture modifications, training procedure, or validation of this encoder. Also, the notation in Eq. (2) defines R as a set of three reference images while using l for the number of reference images and max over j, leaving the reference count ambiguous. These omissions prevent reproducibility and do not rule out bias introduced by using generated images to train the evaluator.
minor comments (6)
  1. [§1] The phrase "lackly coverage" should be "lacks coverage".
  2. [Table 1] The column header "Evalution" is misspelled as "Evalution".
  3. [Table 1 and References] GenEval is cited as reference [39] in Table 1, but reference [39] is GenAI-Bench; in the text GenEval is correctly cited as [25]. This inconsistency should be corrected.
  4. [Figure 3 caption] The caption includes "the the text rendering results," with a duplicated "the."
  5. [Figure 9 caption] The caption begins "Tha task is to give...", which should be "The task is to give...".
  6. [§4.1, Eq. (1)] The edit distance upper bound φ is introduced after the equation; for readability, define φ before presenting the formula.

Circularity Check

1 steps flagged · score 6.0 of 10

Knowledge & Reasoning score is defined as similarity to GPT-4o-generated reference text while GPT-4o is itself one of the scored models, making the reasoning ranking self-referential by construction.

  1. self definitional [Section 4.1 (Metrics), Knowledge and Reasoning paragraph; Tables 2 and 3]
    "GPT-4o is responsible for generating the textual reasoning answers, which serve as the core reference for evaluation. LLM2CLIP then measures the alignment between text and image by calculating the cosine similarity between the GPT-4o-generated answer and the corresponding generated image."

    The Knowledge & Reasoning score is defined as the cosine similarity between an image and a text answer that is itself generated by GPT-4o, a model that also appears as one of the evaluated systems in Tables 2 and 3. The reference standard for 'reasoning' is therefore not an externally verified ground truth but GPT-4o's own textual paraphrase of the expected answer. The metric consequently measures how closely each generated image matches GPT-4o's textual expression, so any model whose output distribution resembles GPT-4o's is rewarded by construction. This is not a speculative bias: GPT-4o obtains the highest Reasoning score on both OneIG-Bench-EN (0.345) and OneIG-Bench-ZH (0.300).

full rationale

OneIG-Bench is a benchmark paper rather than a mathematical derivation, and most of its evaluation dimensions are externally grounded: semantic alignment uses DSG-style question graphs with Qwen2.5-VL answers, text rendering compares extracted strings to the prompt text via edit distance, and diversity uses DreamSim pairwise similarity. These components are self-contained and do not reduce to their own inputs. The one load-bearing circular step is the Knowledge & Reasoning metric in Section 4.1. The paper defines the reasoning reference as GPT-4o-generated text and then scores all models, including GPT-4o image generation, by LLM2CLIP cosine similarity to that text. Since the ground truth is authored by the very model family being ranked, the reasoning score is self-definitional: the quantity the benchmark claims to measure (reasoning quality) is operationally defined as agreement with GPT-4o's output distribution. The paper's own Limitation section concedes that reasoning evaluation is novel and that the only support is an undocumented claim of alignment with human evaluations, which cannot break the circularity because no human study is presented. The style sub-score also uses an encoder fine-tuned on images from CSGO, a method co-authored by one of this paper's authors, but that encoder is averaged with the external CSD metric and is not used as a ground-truth reference, so it does not constitute a derivation-level circularity. Overall, the benchmark has substantial independent content, but its highlighted reasoning dimension reduces by construction to agreement with one of the evaluated models, warranting a partial circularity score of 6.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The evaluation framework rests on several unproven assumptions about the validity of AI-based metrics, especially the reasoning metric that uses GPT-4o as its reference. Free parameters include the text-score normalization constants phi, prompt-length thresholds, Beta-distribution parameters, and the number of style reference images. The OneIG style encoder is an in-house invented component without independent validation.

free parameters (5)
  • phi_EN = 100
    Upper bound for edit distance in the composite text score for OneIG-Bench-EN; chosen by the authors to normalize the text score, no derivation given.
  • phi_ZH = 50
    Upper bound for edit distance in the composite text score for Chinese prompts; set to half the English value because Chinese characters occupy more bytes, a heuristic normalization.
  • short/medium/long word thresholds = 30 and 60 words
    Thresholds for prompt length categories; chosen based on token estimates for CLIP/SigLIP, not derived from data.
  • Beta distribution parameters = 2.37, 2.86
    Used in Algorithm 1 to generate target prompt lengths with a roughly 1:2:1 short/medium/long ratio; selected by hand, not from data.
  • number of reference images per style (l) = 3
    The paper defines R = {R_k^1, R_k^2, R_k^3}, suggesting three reference images per style; this affects style scores and is not justified.
assumptions (6)
  • domain assumption GPT-4o-generated reasoning answers are a valid reference for evaluating image reasoning.
    In Section 4.1, the Knowledge and Reasoning metric uses GPT-4o to generate textual answers, then LLM2CLIP measures similarity to the image. This assumes GPT-4o's answers are a neutral ground truth, which is questionable when GPT-4o is also an evaluated model.
  • domain assumption LLM2CLIP cosine similarity captures reasoning quality in images.
    The reasoning score is the cosine similarity between the GPT-4o answer embedding and the image embedding. This assumes that embedding alignment corresponds to reasoning correctness, which is not validated.
  • domain assumption The DSG question graph generated by GPT-4o provides a reliable semantic-alignment evaluation.
    Alignment scores depend on questions generated by GPT-4o rather than human annotation; this assumes GPT-4o's questions cover the prompt's salient information without bias.
  • domain assumption The style reference images and the two encoders (CSD and OneIG style encoder) adequately capture the target style.
    The style score is the average of cosine similarities from two encoders against curated reference images. This assumes the reference images and encoders are representative of each style.
  • domain assumption The prompt collection and LLM-rewriting pipeline yields diverse, unbiased prompts.
    Prompts are collected from internet/user inputs, clustered, deduplicated, rewritten by GPT-4o, and manually filtered. This assumes these steps do not introduce systematic bias against certain models or styles.
  • ad hoc to paper The word-length ratio 1:2:1 and the chosen thresholds are appropriate for evaluating models at different prompt complexities.
    The authors force a specific length distribution via Beta sampling; this is a design choice that shapes the benchmark, not a standard established elsewhere.
invented entities (1)
  • OneIG style image encoder
    purpose: Style evaluation; a CLIP model fine-tuned on images generated by CSGO to produce style embeddings.
    The encoder is introduced by this paper and is not independently validated. It is trained on images from CSGO, a model co-authored by a member of this team, so the style scores partly depend on an in-house component.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation." pith.science (2026). https://pith.science/paper/STWCDIS6

@misc{pith2026250607977,
  author       = {Pith},
  title        = {Pith review of: OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STWCDIS6}},
  note         = {Machine review of arXiv:2506.07977}
}
read the original abstract

Text-to-image (T2I) models have garnered significant attention for generating high-quality images aligned with text prompts. However, rapid T2I model advancements reveal limitations in early benchmarks, lacking comprehensive evaluations, for example, the evaluation on reasoning, text rendering and style. Notably, recent state-of-the-art models, with their rich knowledge modeling capabilities, show promising results on the image generation problems requiring strong reasoning ability, yet existing evaluation systems have not adequately addressed this frontier. To systematically address these gaps, we introduce OneIG-Bench, a meticulously designed comprehensive benchmark framework for fine-grained evaluation of T2I models across multiple dimensions, including prompt-image alignment, text rendering precision, reasoning-generated content, stylization, and diversity. By structuring the evaluation, this benchmark enables in-depth analysis of model performance, helping researchers and practitioners pinpoint strengths and bottlenecks in the full pipeline of image generation. Specifically, OneIG-Bench enables flexible evaluation by allowing users to focus on a particular evaluation subset. Instead of generating images for the entire set of prompts, users can generate images only for the prompts associated with the selected dimension and complete the corresponding evaluation accordingly. Our codebase and dataset are now publicly available to facilitate reproducible evaluation studies and cross-model comparisons within the T2I research community.

Figures

Figures reproduced from arXiv: 2506.07977 by the authors.

Figure 1
Figure 1. Overview of OneIG-Bench. OneIG-Bench comprises six core categories, each designed to evaluate targeted capabilities across distinct generative dimensions, with approximately 200 carefully curated prompts per category to ensure comprehensive coverage of diverse scenarios. in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The construction pipeline of OneIG-Bench. The construction pipeline comprises four methodical steps to derive our final assessment prompts, ensuring the diversity and comprehensiveness of the benchmark. In the first step, we curated prompts and generation scenes by filtering publicly accessible internet data, user inputs, and some established datasets, thereby ensuring that the benchmark focuses on content aligned w… view at source ↗
Figure 3
Figure 3. An illustration of the generation results and the corresponding scores on OneIG-Bench-ZH. The first row shows the alignment results and the second row shows the the text rendering results. And the evaluation scores are displayed in the upper left corner of the image. 4.2.1 Semantic Alignment and Diversity In the alignment dimension shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: An illustration of the generation results and the corresponding scores. The first row shows the text rendering results and the second row shows the the knowledge and reasoning results. And the evaluation scores are displayed in the upper left corner of the image [PITH…
Figure 5
Figure 5. Figure 5: Word Count of the Overall Prompts of OneIG-Bench. The word count distribution of OneIG￾Bench’s prompts ranges from 0 to 200. Short Middle Long Prompt Length 0.00 0.25 0.50 0.75 1.00 Proportion Portrait General Object Anime & Stylization Text Rendering Knowledge & Reaso…
Figure 7
Figure 7. Figure 7: Visualization results of SOTA methods. Alignment of Imagen4 [17], GPT-4o [45], Imagen3 [36], HiDream-I1-Full [29] and Kolors 2.0 [64] are evaluated respectively. Row 1 corresponds to tag/phrase prompt: The variation in the scores of the visual samples are mainly influe…
Figure 8
Figure 8. Figure 8: Visualization results of SOTA methods. Text of Seedream 3.0 [24], GPT-4o [45], Imagen4 [17], Recraft V3 [66], HiDream-I1-Full [29] are evaluated respectively. Row 1, 2 correspond to short prompts. Row 3, 4 correspond to middle prompts. Row 5, 6 correspond to long promp…
Figure 9
Figure 9. Figure 9: Visualization results of SOTA methods. Reasoning of GPT-4o [45], Imagen4 [17], Recraft V3 [66], HiDream-I1-Full [29] and Imagen3 [36] are evaluated respectively. Row 1 aims to illustrate how coral reefs are formed, highlighting key steps such as the growth of coral pol…
Figure 10
Figure 10. Figure 10: Visualization results of SOTA methods. Diversity of Stable Diffusion 1.5 [55], Janus-Pro [13], Kolors 2.0 [64], Stable Diffusion XL [46] and Seedream 3.0 [24] are evaluated respectively. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Visualization results of SOTA methods. Traditional styles of GPT-4o [45], Stable Diffusion 1.5 [55], SANA-1.5 4.8B (PAG) and 1.6B (PAG) [73], and Recraft V3 [66] are evaluated respectively. The styles are pointillism and minimalism. PROMPT: A muscular man carries an u…
Figure 12
Figure 12. Figure 12: Visualization results of SOTA methods. Media styles of GPT-4o [45], Imagen4 [17], Seedream 3.0 [24], Recraft V3[66] and SANA-1.5 4.8B (PAG) [73] are evaluated respectively. The styles are pencil sketch and stone sculpture. PROMPT: A blonde woman with long, flowing hai…
Figure 13
Figure 13. Figure 13: Visualization results of SOTA methods. Anime styles of Seedream 3.0 [24], BAGEL+CoT [18], Imagen4 [17], BAGEL [18] and Janus-Pro [13] are evaluated respectively The styles are cyberpunk and 3d rendering. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Parallel Decoding Distillation for Fast Image and Video Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A trajectory-based distillation method trains a student to predict multiple mean velocities per network evaluation, enabling 4-8 step generation with competitive quality and improved diversity.

  2. Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    cs.CV 2025-11 unverdicted novelty 6.0 of 10

    Z-Image is an efficient 6B-parameter foundation model for image generation that rivals larger commercial systems in photorealism and bilingual text rendering through a new single-stream diffusion transformer and strea...

  3. Interleaving Reasoning for Better Text-to-Image Generation

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A text-image-text-image multi-turn pipeline improves text-to-image generation, gaining up to 8 points over the base model on several benchmarks while adding an image-conditioned reflection step.

  4. X-Omni: Reinforcement Learning Makes Discrete Autoregressive Image Generative Models Great Again

    cs.CV 2025-07 conditional novelty 6.0 of 10

    GRPO reinforcement learning applied to a discrete autoregressive image generator with a diffusion decoder improves instruction following, image quality, and long-text rendering in a unified multimodal model.

  5. Boogu-Image-0.1: Boosting Open Agentic Multimodal Generation via Understanding under a Minimal Budget

    cs.CV 2026-07 conditional novelty 4.0 of 10

    An open-source image-generation family shows that agentic prompt rewriting and a stronger text encoder can lift quality to near closed-source levels with only 208.62M images and about $400K of training compute.

Reference graph

Works this paper leans on

79 extracted references · 22 canonical work pages · cited by 5 Pith papers

  1. [1]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425–2433, 2015

    2015

  2. [2]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023

    arXiv 2023

  3. [4]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. a...

    arXiv 2025

  4. [5]

    Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models

    Eslam Mohamed Bakr, Pengzhan Sun, Xiaoqian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20041–20053, 2023

    work page 2023

  5. [6]

    The official api of flux-1.dev

    black-forest labs. The official api of flux-1.dev. https://api.us1.bfl.ai/scalar#tag/tasks/POST /v1/flux-dev, 2024

    work page 2024

  6. [7]

    Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG) , 42(4):1–10, 2023

    work page 2023

  7. [8]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset. arXiv preprint arXiv:2505.09568, 2025

    arXiv 2025

  8. [9]

    Pixart-sigma: 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-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024

    work page 2024

Show all 79 references
  1. [10]

    Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023

    arXiv

  2. [11]

    Mmdetection: Open mmlab detection toolbox and benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019

    1906 arXiv

  3. [12]

    Generative pretraining from pixels

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691–1703. PMLR, 2020

    2020

  4. [13]

    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

    2025 arXiv

  5. [14]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vis...

    2024

  6. [15]

    Mask2former for video instance segmentation

    Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021

    2021 arXiv

  7. [16]

    Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation

    Jaemin Cho, Yushi Hu, Jason Baldridge, Roopal Garg, Peter Anderson, Ranjay Krishna, Mohit Bansal, Jordi Pont-Tuset, and Su Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation. In ICLR, 2024

    2024

  8. [17]

    Google deepmind Imagen4 team. Imagen4. https://storage.googleapis.com/deepmind-media /Model-Cards/Imagen-4-Model-Card.pdf , 2025

    2025

  9. [18]

    Emerging properties in unified multimodal pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025

    2025 arXiv

  10. [19]

    Scaling rectified flow transformers 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 transformers for high-resolution image synthesis. In Forty-first international conference on machine ...

    2024

  11. [20]

    Training-free structured diffusion guidance for compositional text-to-image synthesis

    Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022

    2022 arXiv

  12. [21]

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023

    2023

  13. [22]

    Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546, 2024

    Xingyu Fu, Muyu He, Yujie Lu, William Yang Wang, and Dan Roth. Commonsense-t2i challenge: Can text-to-image generation models understand commonsense? arXiv preprint arXiv:2406.07546, 2024

    2024 arXiv

  14. [23]

    Distilling diversity and control in diffusion models

    Rohit Gandikota and David Bau. Distilling diversity and control in diffusion models. arXiv preprint arXiv:2503.10637, 2025

    2025

  15. [24]

    Seedream 3.0 technical report

    Yu Gao, Lixue Gong, Qiushan Guo, Xiaoxia Hou, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, et al. Seedream 3.0 technical report. arXiv preprint arXiv:2504.11346, 2025

    2025 arXiv

  16. [25]

    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

  17. [26]

    Generative adversarial networks.Communications of the ACM, 63(11):139– 144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139– 144, 2020

    2020

  18. [27]

    Evalmuse-40k: A reliable and fine-grained benchmark with comprehensive human annotations for text-to-image generation model evaluation, 2024

    Shuhao Han, Haotian Fan, Jiachen Fu, Liang Li, Tao Li, Junhui Cui, Yunqiu Wang, Yang Tai, Jingwei Sun, Chunle Guo, and Chongyi Li. Evalmuse-40k: A reliable and fine-grained benchmark with comprehensive human annotations for text-to-image generation model evaluation, 2024

    2024

  19. [28]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

    2017

  20. [29]

    Hidream-i1

    HiDream-ai. Hidream-i1. https://github.com/HiDream-ai/HiDream-I1 , 2025

    2025

  21. [30]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  22. [31]

    Ella: Equip diffusion models with llm for enhanced semantic alignment

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

    2024 arXiv

  23. [32]

    Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering

    Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages ...

    2023

  24. [33]

    T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation

    Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  25. [34]

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

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

    2023

  26. [35]

    Llm2clip: Powerful language model unlock richer visual representation

    Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024

    2024

  27. [36]

    Imagen 3, 2024

    Imagen-Team-Google. Imagen 3, 2024

    2024

  28. [37]

    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

  29. [38]

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

    2024

  30. [39]

    Genai-bench: Evaluating and improving compositional text-to-visual generation

    Baiqi Li, Zhiqiu Lin, Deepak Pathak, Jiayao Li, Yixin Fei, Kewen Wu, Tiffany Ling, Xide Xia, Pengchuan Zhang, Graham Neubig, et al. Genai-bench: Evaluating and improving compositional text-to-visual generation. arXiv preprint arXiv:2406.13743, 2024

    2024 arXiv

  31. [40]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

    2024

  32. [41]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888–12900. PMLR, 2022

    2022

  33. [42]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, pages 23–40. Springer, 2024. 14

    2024

  34. [43]

    Wise: A world knowledge-informed semantic evaluation for text-to-image generation

    Yuwei Niu, Munan Ning, Mengren Zheng, Bin Lin, Peng Jin, Jiaqi Liao, Kunpeng Ning, Bin Zhu, and Li Yuan. Wise: A world knowledge-informed semantic evaluation for text-to-image generation. arXiv preprint arXiv:2503.07265, 2025

    2025 arXiv

  35. [44]

    Gpt-4o system card

    OpenAI. Gpt-4o system card. https://openai.com/index/gpt-4o-system-card/ , 2024

    2024

  36. [45]

    Introducing 4o image generation

    OpenAI. Introducing 4o image generation. https://openai.com/index/introducing-4o-image -generation/, 2025

    2025

  37. [46]

    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

  38. [47]

    Lumina-image 2.0: A unified and efficient image generative framework, 2025

    Qi Qin, Le Zhuo, Yi Xin, Ruoyi Du, Zhen Li, Bin Fu, Yiting Lu, Xinyue Li, Dongyang Liu, Xiangyang Zhu, Will Beddow, Erwann Millon, Wenhai Wang Victor Perez, Yu Qiao, Bo Zhang, Xiaohong Liu, Hongsheng Li, Chang Xu, and Peng Gao. Lumina-image 2.0: A unified and efficient image g...

    2025

  39. [48]

    Phybench: Holistic evaluation of physical perception and reasoning in large language models

    Shi Qiu, Shaoyang Guo, Zhuo-Yang Song, Yunbo Sun, Zeyu Cai, Jiashen Wei, Tianyu Luo, Yixuan Yin, Haoxu Zhang, Yi Hu, et al. Phybench: Holistic evaluation of physical perception and reasoning in large language models. arXiv preprint arXiv:2504.16074, 2025

    2025 arXiv

  40. [50]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

    2021

  41. [51]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  42. [52]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020

    2020

  43. [53]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    2022 arXiv

  44. [54]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning , pages 8821–8831. Pmlr, 2021

    2021

  45. [55]

    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 (CVPR), pages 10684–10695, June 2022

    2022

  46. [56]

    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 informatio...

    2022

  47. [57]

    Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis

    Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105–30118. PMLR, 2023

    2023

  48. [58]

    Measuring style similarity in diffusion models

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024

    2024 arXiv

  49. [59]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  50. [60]

    stable-diffusion-3.5-large

    Stability-AI. stable-diffusion-3.5-large. https://github.com/Stability-AI/sd3.5, 2024

    2024

  51. [61]

    Autore- gressive model beats diffusion: Llama for scalable image generation

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

    2024 arXiv

  52. [62]

    Evalalign: Evaluating text-to-image models through precision alignment of multimodal large models with supervised fine-tuning to human annotations

    Zhiyu Tan, Xiaomeng Yang, Luozheng Qin, Mengping Yang, Cheng Zhang, and Hao Li. Evalalign: Evaluating text-to-image models through precision alignment of multimodal large models with supervised fine-tuning to human annotations. arXiv e-prints, pages arXiv–2406, 2024. 15

    2024

  53. [63]

    Chameleon: Mixed-modal early-fusion foundation models

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024

    2024 arXiv

  54. [64]

    Kolors2.0

    Kuaishou Kolors team. Kolors2.0. https://app.klingai.com/cn/, 2025

    2025

  55. [65]

    How to create sota image generation with text recrafts ml team insights

    Recraft team. How to create sota image generation with text recrafts ml team insights. https://www.re craft.ai/blog/how-to-create-sota-image-generation-with-text-recrafts-ml-team-i nsights, 2024

    2024

  56. [66]

    Recraft v3

    Recraft team. Recraft v3. https://www.recraft.ai/blog/recraft-introduces-a-revolutio nary-ai-model-that-thinks-in-design-language?utm_source=ai-bot.cn , 2024

    2024

  57. [67]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  58. [68]

    Emu3: Next-token prediction is all you need

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024

    2024 arXiv

  59. [69]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004

    2004

  60. [70]

    Revisiting text-to-image evaluation with gecko: On metrics, prompts, and human ratings

    Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kaji´c, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Pinelopi Papalampidi, Ira Ktena, Chris Knutsen, et al. Revisiting text-to-image evaluation with gecko: On metrics, prompts, and human ratings. arXiv preprint arXiv:2404...

    2024 arXiv

  61. [71]

    Omnigen2: Exploration to advanced multimodal generation

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, Ze Liu, Ziyi Xia, Chaofan Li, Haoge Deng, Jiahao Wang, Kun Luo, Bo Zhang, Defu Lian, Xinlong Wang, Zhongyuan Wang, Tiejun Huang, and Zheng Liu. Omnigen2: Ex...

    2025 arXiv

  62. [72]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341, 2023

    2023 arXiv

  63. [73]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer, 2025

    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, 2025

    2025

  64. [74]

    Show-o: One single transformer 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 transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024

    2024 arXiv

  65. [75]

    Show-o2: Improved native unified multimodal models

    Jinheng Xie, Zhenheng Yang, and Mike Zheng Shou. Show-o2: Improved native unified multimodal models. arXiv preprint arXiv:2506.15564, 2025

    2025 arXiv

  66. [76]

    Csgo: Content-style composition in text-to-image generation

    Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024

    2024 arXiv

  67. [77]

    Scaling autoregressive models for content-rich text-to-image generation

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022

    2022 arXiv

  68. [78]

    Cogview4

    THUKEG Z.ai. Cogview4. https://github.com/THUDM/CogView4, 2025

    2025

  69. [79]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

    2023

  70. [80]

    Worldgenbench: A world-knowledge-integrated benchmark for reasoning-driven text-to-image generation, 2025

    Daoan Zhang, Che Jiang, Ruoshi Xu, Biaoxiang Chen, Zijian Jin, Yutian Lu, Jianguo Zhang, Liang Yong, Jiebo Luo, and Shengda Luo. Worldgenbench: A world-knowledge-integrated benchmark for reasoning-driven text-to-image generation, 2025. 16 A Appendix A.1 Word Count Distribution...

    2025

  71. [81]

    K & R" is the abbreviation for

    "K & R" is the abbreviation for "Knowledge & Reasoning". Category Avg Short Middle Long Portrait 56.4 0.184 0.443 0.373 General Object 46.5 0.330 0.422 0.248 Anime & Stylization 50.6 0.212 0.522 0.265 Text Rendering 51.2 0.275 0.475 0.250 Knowledge & Reasoning 20.5 0.960 0.018...

    2024

Pith tools

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