Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

Skywork UniPic: Unified Autoregressive Modeling for Visual Understanding and Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 1.5B unified autoregressive model matches or beats much larger rivals on image generation and editing.

desk verdict A genuine 1.5B unified model with competitive generation numbers, but the editing 'leading' claim is contradicted by its own tables and the lack of decontamination analysis leaves the headline scores unsecured. read the letter →

arxiv 2508.03320 v1 pith:3UJQ4IUI submitted 2025-08-05 cs.CV

Peiyu Wang , Yi Peng , Yimeng Gan , Liang Hu , Tianyidan Xie , Xiaokun Wang , Yichen Wei , Chuanxin Tang
show 7 more authors
Bo Zhu Changshi Li Hongyang Wei Eric Li Xuchen Song Yang Liu Yahui Zhou
This is my paper
classification cs.CV
keywords unifiedmultimodalmodelautoregressiveimagegenerationeditingdecoupledvisualencodingmaskedSigLIP2reward-modeldatacurationparameterefficiency
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 is trying to establish that a single 1.5-billion-parameter autoregressive model can handle image understanding, text-to-image generation, and image editing at the same time, without task-specific adapters or inter-module connectors. The proposed design, Skywork UniPic, pairs a masked-autoregressive encoder-decoder for generation with a SigLIP2 encoder for understanding, and routes both through one shared language-model backbone trained under a unified objective. The authors report state-of-the-art unified-model results on GenEval (0.86), DPG-Bench (85.5), GEditBench-EN (5.83), and ImgEdit-Bench (3.49), while generating 1024-by-1024 images with under 15 GB of GPU memory. If these results hold, a compact model can deliver what currently requires 7B to 19B parameters, which would make deployable multimodal AI practical on commodity hardware.

What carries the argument

The load-bearing mechanism is the decoupled encoding strategy inside a unified autoregressive framework. A Masked Autoregressive (MAR) encoder-decoder, a pixel-space generative model without vector quantization, handles synthesis; a SigLIP2 visual encoder, a contrastively trained vision-language encoder, handles understanding; and separate MLP projections feed both into a shared 1.5B language-model backbone that emits either image tokens or text under one multi-task loss. A four-stage curriculum (MAR pretraining, MAR-LLM alignment, joint optimization, supervised fine-tuning) progressively scales resolution from 256x256 to 1024x1024 while unfreezing parameters, and two task-specific reward models filter training samples, so that the data quality and training schedule carry much of the claimed efficiency.

What would settle it

Sample the released training corpus and run near-duplicate retrieval for the exact prompts used in GenEval, DPG-Bench, GEditBench-EN, and ImgEdit-Bench; if any nontrivial fraction of evaluation prompts or their reference images appears, the benchmark gains could be memorization. Alternatively, evaluate the released model on a freshly written set of composition and editing prompts in the same style and check whether the 0.86, 85.5, 5.83, and 3.49 scores hold on data the authors never saw.

Watch

Extended reading notes

Core claim

The central claim is that the apparent conflict between pixel-level fidelity and semantic comprehension in unified visual models can be resolved by decoupling the visual encoders rather than by scaling up parameters. Skywork UniPic uses a Masked Autoregressive (MAR) encoder-decoder that operates on pixel-space latents for generation and a SigLIP2 encoder that produces semantic features for understanding, with both feeding a single autoregressive language-model backbone through separate projection layers. All tasks are trained together with one weighted objective, and the paper argues that this shared decoder lets generation improve visual detail modeling and lets understanding guide coherent editing. On the four benchmarks the model reports 0.86 on GenEval, 85.5 on DPG-Bench, 5.83 on GEditBench-EN, and 3.49 on ImgEdit-Bench, which the authors attribute to decoupled encoding, progressive resolution-aware training from 256x256 to 1024x1024, and hundred-million-scale reward-filtered data rather than to model size.

Load-bearing premise

The reported state-of-the-art numbers are meaningful only if the 130M-sample training corpus does not contain the GenEval, DPG-Bench, GEditBench-EN, or ImgEdit-Bench prompts or their outputs, and the paper gives no decontamination analysis to establish that.

Editorial extensions

If this is right

  • Unified multimodal capability can be delivered at 1.5B scale: competitive generation and editing no longer require a 7B-19B model, so the same architecture class becomes feasible on a single consumer GPU.
  • Decoupling the generation encoder from the understanding encoder while keeping one shared language backbone is a workable template for other unified visual tasks, such as in-context editing or interleaved image-text dialogue.
  • A progressive, resolution-aware curriculum produces staged capability emergence, with generation appearing before editing, and that ordering can be used to schedule training stages deliberately.
  • Reward-model-filtered data curation at the hundred-million scale can substitute for raw data volume, making data quality rather than dataset size the lever for instruction-following and edit precision.
  • Releasing weights and code lets the benchmark numbers be checked and the model be used in resource-constrained settings, making the efficiency claim directly verifiable.

Reading between the lines

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

  • The two-encoder-plus-shared-backbone layout is not tied to the specific backbone this paper uses, so the same recipe could plausibly be lifted to other language models or higher resolutions; the paper only demonstrates one configuration, so this is an extrapolation.
  • Because the paper reports no decontamination analysis, the headline benchmark numbers are best read as upper bounds until independent evaluation on fresh, similar prompts confirms that the gains come from generalization rather than training-set overlap.
  • If the observed resolution-dip-then-recovery pattern is general, scaling to 2048x2048 would be expected to cause a temporary quality drop followed by a higher plateau, which is a testable prediction for follow-up training runs.
  • The staged emergence of editing after generation suggests a practical curriculum rule: introduce editing data only after generation is stable, and do not judge a model's editing ability from early checkpoints.
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. Skywork UniPic proposes a unified autoregressive model that combines a masked autoregressive (MAR) encoder/decoder for image generation, a SigLIP2 encoder for image understanding, and a shared Qwen2.5-1.5B language model. The training recipe is a four-stage curriculum with progressive resolution scaling and reward-model-filtered data. The paper reports GenEval 0.86, DPG-Bench 85.5, GEditBench-EN 5.83, ImgEdit-Bench 3.49, and memory usage under 15 GB on an RTX 4090, and it releases code and weights.

Significance. If the reported scores survive scrutiny, the paper would make a useful point: a compact autoregressive model with a decoupled generation/understanding encoder design can reach the top tier of unified models on text-to-image benchmarks while remaining deployable on commodity hardware. The decoupled MAR/SigLIP2 encoding strategy is a reasonable response to the known tension between pixel-level fidelity and semantic representation, and the staged training schedule with reward-model-filtered data is concrete and reproducible. The open release of code and weights is a clear strength because the claims can be independently checked. However, the paper's own tables and evaluation sections undercut several of the headline claims as currently worded, and the benchmark scores lack decontamination evidence, so the stated state-of-the-art conclusions are not yet supported.

major comments (5)
  1. [Abstract; §4.3, Tables 4 and 5] The Abstract and Introduction claim that the model is 'leading among unified models on editing tasks' and establishes 'new state-of-the-art results across critical benchmarks.' This is contradicted by the paper's own tables: on GEditBench-EN (Table 4), Skywork UniPic's 5.83 is below OmniGen2 (6.41), BAGEL (6.52), and Ovis-U1 (6.42); on ImgEdit-Bench (Table 5), its 3.49 is below Ovis-U1 (4.00). Even the GenEval number is not the top among unified models, since Table 2 lists Ovis-U1 at 0.89 and BAGEL with prompt rewriting at 0.88. The body text in Section 4.3 is appropriately hedged with 'among the top-tier unified models,' but the abstract and introduction are not. These claims must be revised to match the reported data, or the authors need to state a different comparison protocol that justifies the wording.
  2. [§4.1, §4.2, §4.3, §5] The paper's stated goal is unification of 'visual understanding, text-to-image generation, and image editing,' and Section 4.1 promises evaluation of 'image understanding, text-to-image generation, and image editing.' Yet Sections 4.2 and 4.3 report only generation and editing benchmarks. The only allusion to understanding results is the qualitative remark in Section 5 about tracking 'standard VLM benchmarks as a proxy,' with no benchmark names, no scores, and no table. Since 'visual understanding' appears in the title and in the contribution list, the central unified-capability claim is currently unevaluated. Please add a quantitative understanding benchmark table (for example, MME, MMBench, or a comparable suite) or explicitly narrow the claimed scope of the model.
  3. [§3.3, §4.1] No decontamination analysis is reported between the training corpus and the evaluation benchmarks, which is a load-bearing gap for every benchmark-based claim. Section 3.3 states that the editing reward model is trained on HumanEdit, UltraEdit, and SuperEdit-40K and that curated training data is filtered with reward models; Section 4.1 evaluates on GEditBench-EN, which originates from the Step1X-Edit pipeline, and on ImgEdit-Bench. Without exact and near-duplicate overlap analysis at both prompt and image levels, it remains possible that the editing scores reflect memorization of benchmark items present in the SFT or reward-model training data. The same concern applies to GenEval prompts, which are short and could appear in web-scale caption collections. Please provide a decontamination analysis or otherwise bound the overlap between the training corpora and all four benchmarks.
  4. [Abstract; §3.1, Table 1, §4.1] The parameter-count claims are internally inconsistent, which affects the paper's headline efficiency argument. The abstract calls Skywork UniPic a '1.5 billion-parameter autoregressive model,' and Section 4.1 says '1.5B activated parameters.' However, Section 3.1 states that the MAR encoder/decoder pair contains approximately 1B parameters with 20 layers each, and Table 1 lists an 800M MAR module combined with a 1.5B language model. The total parameter count is therefore approximately 2.3-2.5B, not 1.5B. The comparisons to 14B-19B unified models depend on whether total, backbone-only, or per-task activated parameters are being counted. Please report total, activated, and per-task parameter counts consistently and state which count is used in each efficiency comparison.
  5. [§4.1, Tables 2-5] Section 4.1 states that all scores come from single evaluation runs without reranking or multi-sampling. This is transparent, but the paper's 'surpassing' and 'new record' claims often rest on very small margins: DPG-Bench 85.5 versus BAGEL 85.07, and GenEval 0.86 versus Ovis-U1 0.89. Given that the evaluation protocol involves generated samples, single-run differences of this size are not clearly meaningful. Please report standard deviations across repeated runs or seeds, or at least provide the official benchmark's known variability, for the specific comparisons used to support state-of-the-art claims.
minor comments (6)
  1. [§3.2] The equations for L_Gen, L_Und, and L_Total are garbled or incomplete in the manuscript and need to be rewritten with proper mathematical notation so that the multi-task objective is actually specified.
  2. [§3.1] The footnote markers for 'MAR-Huge*' and 'SigLIP2-so400m-patch16-512†' appear in the text, but the footnote contents are missing. If these notes contain implementation details or license information, they should be included.
  3. [References] Reference [55] is listed as '1.58-bit flux,' but the text cites it for FLUX.1-dev. Please verify that this citation is correct, since the cited title does not match the model name used in the comparison tables.
  4. [Figure 3, §5] Figure 3 mentions 'in-context tasks' and Section 5 refers to 'standard VLM benchmarks,' but neither is defined or accompanied by quantitative results. Please either define these terms and give scores, or remove them from the narrative.
  5. [§4.4, Figures 4-6] The qualitative figures do not describe how examples were selected. A short statement on the selection protocol would help rule out apparent cherry-picking in success and failure cases.
  6. [Table 1] The 'Training samples' row lists 130M for PT, Alignment, and CT, which is ambiguous: is the same 130M corpus reused across stages, or is it re-sampled? Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all headline results are evaluated against external benchmarks and no reported score reduces to a fitted parameter or self-citation.

full rationale

The paper's central claims—GenEval 0.86, DPG-Bench 85.5, GEditBench-EN 5.83, ImgEdit-Bench 3.49—are empirical scores computed by official benchmark scripts on external benchmarks (Section 4.1). The model's internal components (MAR encoder-decoder, SigLIP2 encoder, Qwen2.5 backbone) and reward models are used for training and data filtering, not for evaluating the reported results, so no equation in the paper reduces a prediction to a training-time fit. The generation and understanding losses are standard multi-task objectives, and the reward models (Skywork-ImgReward, Skywork-EditReward) are trained on preference and editing datasets separate from the evaluation benchmarks; none of their outputs feed the reported GenEval, DPG-Bench, GEdit-Bench, or ImgEdit-Bench scores. The paper cites prior work such as Harmon, MAR, SigLIP2, and Step1X-Edit, but these citations supply architecture components and benchmark definitions, not the numerical results being claimed, and no load-bearing premise relies on an unverified self-citation. Separate concerns—the absence of a decontamination analysis for the 130M-sample corpus and the abstract's 'leading among unified models on editing tasks' being contradicted by Tables 4 and 5—are external-validity and reporting-accuracy issues, not circularity. Accordingly, the derivation is self-contained with respect to circularity concerns.

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

The central claims are empirical system results, so the ledger records the hand-chosen training and evaluation choices that the benchmark numbers rest on, plus the unstated background assumptions about data contamination and metric validity. The model's learned weights are not enumerated individually; they are the fitted object of the paper rather than a separate free parameter.

free parameters (4)
  • Evaluation decoding configuration = 64 sampling steps, CFG scale 3, 1024x1024 resolution
    Chosen by hand; all reported generation and editing scores depend on these settings, and no sensitivity analysis is provided.
  • Reward filtering threshold = 0.9
    Training samples with reward scores below 0.9 are discarded. This hand-chosen threshold defines the curated training set and therefore influences all downstream benchmark results.
  • Multi-task loss weights (U:G:E) = 0:1:0 in pretraining; 0.01:1:1 in CT and SFT
    The balance between understanding, generation, and editing losses is scheduled by hand across stages and directly shapes the final model behavior.
  • Resolution training schedule = 256x256 to 512x512 to 1024x1024
    The progressive resolution schedule is a hand-designed curriculum; the paper reports that resolution increases unlock higher performance ceilings.
assumptions (4)
  • domain assumption A single shared LLM can decode both MAR and SigLIP2 visual features without destructive interference.
    The whole architecture rests on the claim that both encoders can feed one Qwen2.5 backbone and transfer knowledge across tasks. This is stated in Section 3.1 as the core design rationale.
  • domain assumption The 130M-sample training corpus does not overlap with GenEval, DPG-Bench, GEdit-Bench, or ImgEdit-Bench.
    No decontamination analysis is provided. Section 3.3 describes data curation but not benchmark overlap checks; if leakage occurred, the reported scores would not measure generalization.
  • domain assumption The reward models give a reliable ordinal quality signal for filtering training data.
    Skywork-ImgReward and Skywork-EditReward are trained on preference datasets and then used to discard samples below a threshold. Their reliability on new data is assumed in Section 3.3.
  • domain assumption The automatic metrics used by the external benchmarks are trustworthy proxies for generation and editing quality.
    DPG-Bench and GEdit-Bench rely on automated scoring. The paper uses official scripts but does not validate the metrics against human judgments in this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Skywork UniPic: Unified Autoregressive Modeling for Visual Understanding and Generation." pith.science (2026). https://pith.science/paper/3UJQ4IUI

@misc{pith2026250803320,
  author       = {Pith},
  title        = {Pith review of: Skywork UniPic: Unified Autoregressive Modeling for Visual Understanding and Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3UJQ4IUI}},
  note         = {Machine review of arXiv:2508.03320}
}
read the original abstract

We introduce Skywork UniPic, a 1.5 billion-parameter autoregressive model that unifies image understanding, text-to-image generation, and image editing within a single architecture-eliminating the need for task-specific adapters or inter-module connectors-and demonstrate that compact multimodal systems can achieve state-of-the-art performance on commodity hardware. Skywork UniPic achieves a GenEval score of 0.86, surpassing most existing unified models; sets a new DPG-Bench complex-generation record of 85.5; attains 5.83 on GEditBench-EN and 3.49 on ImgEdit-Bench for image editing; and generates 1024 x 1024 images with under 15 GB of GPU memory (e.g., RTX 4090). (1) a decoupled encoding strategy that leverages a masked autoregressive encoder for synthesis and a SigLIP2 encoder for understanding, all feeding a shared autoregressive decoder; (2) a progressive, resolution-aware training schedule scaling from 256 x 256 to 1024 x 1024 while dynamically unfreezing parameters to balance capacity and stability; and (3) meticulously curated, 100 million-scale datasets augmented with task-specific reward models to refine generation and editing objectives. By demonstrating that high-fidelity multimodal integration need not incur prohibitive resource demands, Skywork UniPic establishes a practical paradigm for deployable, high-fidelity multimodal AI. Code and weights are publicly available at https://huggingface.co/Skywork/Skywork-UniPic-1.5B.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. DynEval: Holistic Evaluations of T2I Generative Models in the Wild

    cs.CV 2026-07 conditional novelty 6.5 of 10

    DynEval distills a 235B teacher VLM into 2B/4B evaluators via 250K synthetic instruction triplets, yielding higher human correlation than existing T2I metrics while enabling open-set dynamic QA and scene-graph quality checks.

  2. Reconstruction Alignment Improves Unified Multimodal Models

    cs.CV 2025-09 conditional novelty 6.0 of 10

    RECA, a self-supervised post-training objective that conditions unified multimodal models on their own visual understanding embeddings to reconstruct input images, improves text-to-image and editing benchmarks across ...

Reference graph

Works this paper leans on

62 extracted references · 56 canonical work pages · cited by 2 Pith papers

  1. [1]

    Stable diffusion 3 medium: Multimodal diffusion transformer for photorealistic text-to-image generation

    Stability AI. Stable diffusion 3 medium: Multimodal diffusion transformer for photorealistic text-to-image generation. https://stability.ai/news/stable-diffusion-3-medium ,

    work page

  2. [2]

    Deep- speed inference: Enabling efficient inference of transformer models at unprecedented scale,

    Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, Ammar Ahmad Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunji Ruwase, and Yuxiong He. Deep- speed inference: Enabling efficient inference of transformer models at unprecedented scale,

    work page

  3. [3]

    Humanedit: A high-quality human-rewarded dataset for instruction-based image editing,

    Jinbin Bai, Wei Chow, Ling Yang, Xiangtai Li, Juncheng Li, Hanwang Zhang, and Shuicheng Yan. Humanedit: A high-quality human-rewarded dataset for instruction-based image editing,

    work page

  4. [4]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023

    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, 2023. 7

    work page 2023

  5. [5]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 4, 8, 11, 12

    work page 2023

  6. [6]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, Le Xue, Caiming Xiong, and Ran Xu. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset, 2025. 5, 8, 10, 11

    work page 2025

  7. [7]

    Pixart- � : Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- � : Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 8, 11

    work page 2023

  8. [8]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling, 2025

    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, 2025. 4, 8, 10, 11

    work page 2025

Show all 62 references
  1. [9]

    Emerging properties in unified multimodal pretraining, 2025

    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, 2025. 3, 5, 8, 10, 11, 12

    2025

  2. [10]

    Taming transformers for high-resolution image synthesis, 2021

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. 4

    2021

  3. [11]

    Unified autoregressive visual generation and under- standing with continuous tokens

    Lijie Fan, Luming Tang, Siyang Qin, Tianhong Li, Xuan Yang, Siyuan Qiao, Andreas Steiner, Chen Sun, Yuanzhen Li, Tao Zhu, et al. Unified autoregressive visual generation and under- standing with continuous tokens. arXiv preprint arXiv:2503.13436, 2025. 5

    2025 arXiv

  4. [12]

    Geneval: An object-focused framework for evaluating text-to-image alignment, 2023

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment, 2023. 3, 8

    2023

  5. [13]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 4

    2014

  6. [14]

    Gemini 2.0 flash

    Google. Gemini 2.0 flash. https://developers.googleblog.com/en/experiment-with-gemini-20- flash-native-image-generation, 2025. 8, 11

    2025

  7. [15]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 4

    2020

  8. [16]

    Ella: Equip diffusion models with llm for enhanced semantic alignment, 2024

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment, 2024. 3, 8

    2024

  9. [17]

    Anyedit: Edit any knowledge encoded in language models,

    Houcheng Jiang, Junfeng Fang, Ningyu Zhang, Guojun Ma, Mingyang Wan, Xiang Wang, Xiangnan He, and Tat seng Chua. Anyedit: Edit any knowledge encoded in language models,

  10. [18]

    Nova: Generative language models for assembly code with hierarchical attention and contrastive learning, 2025

    Nan Jiang, Chengxiao Wang, Kevin Liu, Xiangzhe Xu, Lin Tan, Xiangyu Zhang, and Petr Babkin. Nova: Generative language models for assembly code with hierarchical attention and contrastive learning, 2025. 8, 10

    2025

  11. [19]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023

    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, 2023. 7

    2023

  12. [20]

    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. 4, 11

    2024

  13. [21]

    Superedit: Rectifying and facilitating supervision for instruction-based image editing, 2025

    Ming Li, Xin Gu, Fan Chen, Xiaoying Xing, Longyin Wen, Chen Chen, and Sijie Zhu. Superedit: Rectifying and facilitating supervision for instruction-based image editing, 2025. 8

    2025

  14. [22]

    Autoregressive image generation without vector quantization, 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. 3, 4, 5, 6

    2024

  15. [23]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Ji...

    2024

  16. [24]

    Uniworld-v1: High-resolution semantic encoders for unified visual understanding and generation, 2025

    Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, Yatian Pang, and Li Yuan. Uniworld-v1: High-resolution semantic encoders for unified visual understanding and generation, 2025. 3, 8, 10, 11, 12

    2025

  17. [25]

    Evaluating text-to-visual generation with image-to-text generation, 2024

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. 8

    2024

  18. [26]

    Step1x-edit: A practical framework for general image editing, 2025

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, Guopeng Li, Yuang Peng, Quan Sun, Jingwei Wu, Yan Cai, Zheng Ge, Ranchen Ming, Lei Xia, Xianfang Zeng, Yibo Zhu, Binxing Jiao, Xiangyu Zhang, Gang Yu, and Dax...

    2025

  19. [27]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. 4

    2022

  20. [28]

    Improving image generation with better captions

    OpenAI. Improving image generation with better captions. https://cdn.openai.com/ papers/dall-e-3.pdf. 4, 8, 11

  21. [29]

    Gpt-4o system card, 2024

    OpenAI. Gpt-4o system card, 2024. 1

    2024

  22. [30]

    OpenAI. Gpt-4o. https://openai.com/index/introducing-4o-image-generation, 2025. 8, 10, 11, 12

    2025

  23. [31]

    Transfer between modalities with metaqueries, 2025

    Xichen Pan, Satya Narayan Shukla, Aashu Singh, Zhuokai Zhao, Shlok Kumar Mishra, Jialiang Wang, Zhiyang Xu, Jiuhai Chen, Kunpeng Li, Felix Juefei-Xu, Ji Hou, and Saining Xie. Transfer between modalities with metaqueries, 2025. 1, 5, 8, 10

    2025

  24. [32]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    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, 2023. 4, 8, 10, 11

    2023

  25. [33]

    Du, Zehuan Yuan, and Xinglong Wu

    Liao Qu, Huichao Zhang, Yiheng Liu, Xu Wang, Yi Jiang, Yiming Gao, Hu Ye, Daniel K. Du, Zehuan Yuan, and Xinglong Wu. Tokenflow: Unified image tokenizer for multimodal understanding and generation, 2024. 8, 10, 11 18

    2024

  26. [34]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025

  27. [35]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3

    2021

  28. [36]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 4, 5, 10

    2022

  29. [37]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. 3, 7

    2024

  30. [38]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021. 4

    2021

  31. [39]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 4

    2024 arXiv

  32. [40]

    Generative multimodal models are in-context learners, 2024

    Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners, 2024. 1, 5

    2024

  33. [41]

    Siglip 2: Multilingual vision- language encoders with improved semantic understanding, localization, and dense features,

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alab- dulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. Siglip 2: Multilingual vision- language ...

  34. [42]

    Neural discrete representation learning, 2018

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning, 2018. 1

    2018

  35. [43]

    Illume: Illuminating your llms to see, draw, and self-enhance, 2024

    Chunwei Wang, Guansong Lu, Junwei Yang, Runhui Huang, Jianhua Han, Lu Hou, Wei Zhang, and Hang Xu. Illume: Illuminating your llms to see, draw, and self-enhance, 2024. 1, 5

    2024

  36. [44]

    Ovis-u1 technical report, 2025

    Guo-Hua Wang, Shanshan Zhao, Xinjie Zhang, Liangfu Cao, Pengxin Zhan, Lunhao Duan, Shiyin Lu, Minghao Fu, Xiaohao Chen, Jianshan Zhao, Yang Li, and Qing-Guo Chen. Ovis-u1 technical report, 2025. 8, 10, 11, 12

    2025

  37. [45]

    Emu3: Next-token prediction is all you need, 2024

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tie...

    2024

  38. [46]

    Janus: Decoupling visual encoding for unified multimodal understanding and generation, 2024

    Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, and Ping Luo. Janus: Decoupling visual encoding for unified multimodal understanding and generation, 2024. 4, 8, 10, 11

    2024

  39. [47]

    Omnigen2: Exploration to advanced multimodal generation, 2025

    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

  40. [48]

    Openuni: A simple baseline for unified multimodal understanding and generation

    Size Wu, Zhonghua Wu, Zerui Gong, Qingyi Tao, Sheng Jin, Qinyue Li, Wei Li, and Chen Change Loy. Openuni: A simple baseline for unified multimodal understanding and generation. arXiv preprint arXiv:2505.23661, 2025. 5 19

    2025 arXiv

  41. [49]

    Harmonizing visual representations for unified multimodal understanding and generation, 2025

    Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Zhonghua Wu, Qingyi Tao, Wentao Liu, Wei Li, and Chen Change Loy. Harmonizing visual representations for unified multimodal understanding and generation, 2025. 4, 5

    2025

  42. [50]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 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, 2023. 7

    2023

  43. [51]

    Vila-u: a unified foundation model integrating visual understanding and generation, 2025

    Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, Song Han, and Yao Lu. Vila-u: a unified foundation model integrating visual understanding and generation, 2025. 1

    2025

  44. [52]

    Omnigen: Unified image generation, 2024

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation, 2024. 8, 10, 11, 12

    2024

  45. [53]

    Show-o: One single transformer to unify multimodal understanding and generation, 2024

    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, 2024. 1, 5, 8, 10, 11

    2024

  46. [54]

    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,

  47. [55]

    1.58-bit flux, 2024

    Chenglin Yang, Celong Liu, Xueqing Deng, Dongwon Kim, Xing Mei, Xiaohui Shen, and Liang-Chieh Chen. 1.58-bit flux, 2024. 8, 10, 11

    2024

  48. [56]

    Imgedit: A unified image editing dataset and benchmark, 2025

    Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. Imgedit: A unified image editing dataset and benchmark, 2025. 3, 8

    2025

  49. [57]

    Sigmoid loss for language image pre-training, 2023

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 3

    2023

  50. [58]

    Magicbrush: A manually annotated dataset for instruction-guided image editing, 2024

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing, 2024. 8, 11, 12

    2024

  51. [59]

    In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer, 2025

    Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer, 2025. 4, 8, 11, 12

    2025

  52. [60]

    Ultraedit: Instruction-based fine-grained image editing at scale, 2024

    Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale, 2024. 4, 8, 12

    2024

  53. [61]

    Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. 1

    2024

  54. [62]

    Lumina-next: Making lumina-t2x stronger and faster with next-dit, 2024

    Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, Xu Luo, Zehan Wang, Kaipeng Zhang, Xiangyang Zhu, Si Liu, Xiangyu Yue, Dingning Liu, Wanli Ouyang, Ziwei Liu, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-next...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.