Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

WeEdit shows that text-centric image editing is a data-and-training problem: its glyph-guided fine-tuning plus multi-objective RL lifts instruction adherence from 3.49 to 7.47 on the bilingual benchmark, topping all open-source baselines an

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 18:20 UTC pith:Y6LMILM7

load-bearing objection Read it for the dataset and HTML pipeline; don't trust the SOTA ranking until benchmark overlap is addressed. the 3 major comments →

arxiv 2603.11593 v2 pith:Y6LMILM7 submitted 2026-03-12 cs.CV

WeEdit: A Dataset, Benchmark and Glyph-Guided Framework for Text-centric Image Editing

Hui Zhang , Juntao Liu , Zongkai Liu , Liqiang Niu , Fandong Meng , Zuxuan Wu , Yu-Gang Jiang This is my paper
classification cs.CV
keywords text-centric image editinginstruction-based image editingglyph-guided fine-tuningreinforcement learning for diffusionmultilingual benchmarkHTML-based data pipelinebenchmark datasetdiffusion models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that current image-editing models fail at text-centric editing—modifying, translating, rearranging, or deleting text inside an image—because they lack specialized training data and training paradigms, not because the underlying diffusion architectures are inadequate. To prove this, the authors build WeEdit, a package with three parts: an HTML-based pipeline that automatically produces 330K source-target image pairs with pixel-exact ground truth and 15-language coverage; two benchmarks (bilingual and multilingual) with eight editing operations; and a two-stage training scheme that first injects a glyph image (rendered target text with bounding boxes) as an explicit spatial prior, then applies reinforcement learning with four separate reward models. On the paper's benchmarks, the full model lifts instruction adherence from 3.49 to 7.47 (bilingual) and from 3.63 to 6.70 (multilingual), surpassing all open-source baselines and most proprietary models, and ranking second only to Gemini-3-Pro-Image. The authors attribute the gains mainly to the synergy between glyph-guided SFT and the multi-objective RL stage, with the largest improvements on translation and rearrangement.

Core claim

The central claim is that text-centric editing failures are a data-and-training problem, not a model-capacity problem. With pixel-perfect paired data from the HTML pipeline, an explicit glyph prior, and multi-objective RL, the base editing model's instruction adherence jumps from 3.49 to 7.47 on the bilingual benchmark and from 3.63 to 6.70 on the multilingual benchmark; the paper shows that removing the glyph prior, the SFT stage, or the reference-image/split-reward design each degrades performance, and that gains are largest on translation and rearrangement tasks, which were near-failures for all baselines.

What carries the argument

The HTML pipeline: a VLM converts a source image to HTML/Tailwind, a lightweight VLM edits or translates extracted text entries, the target strings are backfilled at identical DOM positions, and a headless browser renders both versions—giving pixel-perfect pairs where all non-text pixels are identical by construction. The glyph prior: a VLM performs detect-and-plan, outputting target strings and bounding boxes, which become a white-on-black glyph image; the flow-based diffusion model (LoRA-trained on Qwen-Image-Edit) receives original image, glyph, and instruction. The RL stage: DiffusionNFT-style policy optimization with a composite reward that is a weighted sum of four separate VLM scores—

Load-bearing premise

The load-bearing premise is that the proprietary VLM judge's scores reflect true editing quality and that the benchmark cases are not pulled from the same HTML pipeline that generated WeEdit's training data; if the judge is biased toward models trained on similar data or the benchmark is inadvertently in-distribution, the reported SOTA margins could be an artifact of evaluation rather than genuine editing ability.

What would settle it

Take a random sample of 200 cases from the Bilingual benchmark, have three independent human annotators blind-rank WeEdit, Qwen-Image-Edit-2509, and FLUX.2-dev on instruction adherence, and compare the human ordering with the Gemini-3-Pro ordering; if the judge's ordering is not reproduced, the central SOTA claim is not settled. Separately, compute perceptual hashes of benchmark source images and training source images; high overlap would indicate the benchmark is in-distribution for WeEdit.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Open-source editing models can compete with proprietary systems on text-centric tasks using synthetic paired data and glyph conditioning.
  • The HTML pipeline extends to any language or script whose text can be extracted from structured markup, so multilingual coverage is limited mainly by the availability of source pages.
  • The four-way reward decomposition points to a general recipe for fine-grained editing alignment: separate quality dimensions and anchor quality to a reference image.
  • Benchmark tasks that require world knowledge (e.g., replacing a sports champion name) expose whether an editing model can reason, not just render.
  • Because the RL judge is a VLM, the same pipeline could be iterated with stronger judges to push instruction adherence further.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper does not specify how the benchmark cases were curated; if they were drawn from the same web-page sources as the training images, the reported margins may shrink on natural photos. A perceptual-hash overlap check between benchmark and training images would resolve this.
  • The user study is small and covers only two dimensions; a larger, pre-registered human study across all eight operation types would test whether the VLM-judge ranking matches human preference.
  • The glyph prior is rendered with a single font (Arial) on white-black; real posters and UI designs use diverse fonts and colors, so the method's robustness to typographic variety is still an open question that could be probed by changing the glyph renderer.
  • If the approach transfers, the same detect-and-plan plus glyph conditioning idea could be applied to video text editing, or to editing charts and diagrams where text and layout are tightly coupled.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents WeEdit, a dataset, benchmark, and training framework for text-centric image editing. It introduces a 330K-pair training set built from an HTML-based structured pipeline and an unstructured generate-verify-retry pipeline, and constructs bilingual (Chinese/English) and 15-language multilingual benchmarks covering 8 editing operations. The method fine-tunes Qwen-Image-Edit-2509 with LoRA using a VLM-predicted glyph image as an extra spatial condition, followed by a DiffusionNFT-style RL stage with four VLM-based reward dimensions. On its own benchmark, WeEdit-RL reports overall IA/TC/BP of 7.47/8.19/9.01 on the bilingual split and 6.70/7.10/8.49 on the multilingual split, surpassing all open-source baselines and all proprietary baselines except Gemini-3-Pro-Image.

Significance. If the benchmark is cleanly constructed and the judge is trustworthy, this is a substantial contribution. The structured HTML pipeline is a particular strength: deterministic rendering gives pixel-level identity outside edited text regions, and the multilingual extension is natural. The glyph-conditioned SFT provides an explicit spatial prior, and the multi-objective RL reward is carefully decomposed with logit-weighted scoring. The benchmark also includes a Reasoning split outside the training operation set, which is a useful generalization probe. The main risks are evaluation validity and train/test separation; these affect the central SOTA claim but are addressable in revision.

major comments (3)
  1. [4.3 (cf. 4.1)] The benchmark construction is not described. The structured training data in Sec. 4.1 come from Leopard, The Cauldron, WebSight, etc., through a deterministic HTML renderer. The 2,000 bilingual and 2,000 multilingual test cases are called 'meticulously curated,' but no source corpora, filtering procedure, or exclusion-of-training-sources statement is given. If the test cases are produced from the same source images or the same HTML pipeline, WeEdit is evaluated in-distribution while all 11 open-source baselines are zero-shot, and the reported +3.98 IA over the base model and roughly 3.3-point margin over the best open-source baseline could reflect distribution fit rather than editing ability. Please specify benchmark sources and report overlap statistics (image-level and HTML-level) with the training set.
  2. [4.3, 5.2, Fig. 7] All ranking numbers are produced by Gemini-3-Pro, a proprietary VLM judge, with no validation against human scores, no inter-judge agreement, and no analysis of judge bias. The user study covers only four baselines and does not report the number of evaluators, selection procedure, or significance tests; it cannot validate the full 15-model ranking, especially on Translate/Rearrange/Reasoning where the margins are largest. In addition, Sec. 5.1 gives no inference protocol for baselines (resolution, prompt template, sampling parameters, number of runs), which can materially affect text clarity scores. Please add a human-correlation sample, judge reliability statistics, and a precise evaluation protocol.
  3. [5.2, Table 2] The claim that 'the RL stage provides consistent and targeted improvements' is not fully supported on the multilingual benchmark. Comparing WeEdit-RL with WeEdit-SFT in Table 2, BP decreases from 8.58 to 8.49; Style IA drops from 7.98 to 7.14 and Style BP from 8.90 to 8.62; Add TC and BP also decrease slightly. The conclusion should be qualified or the regressions explained.
minor comments (6)
  1. [6.2, Table 3] The ablation table uses only checkmarks as column headings. Add a legend and explicit row labels so each condition (SFT/Glyph/RL/RI/SRM) is unambiguous.
  2. [Tables 1-2] The captions say 'top-3 highlighted,' but the plain-text rendering shows no highlighting. Ensure the final PDF has visible bold/color marks.
  3. [4.3] The phrase 'meticulously curated' is unsupported. Describe the curation/filtering criteria and any human verification used during benchmark construction.
  4. [Fig. 7] Report the number of human evaluations per cell and, if possible, confidence intervals. Without N, the win/tie/lose percentages cannot be interpreted.
  5. [General] No explicit code or data availability statement appears in the manuscript. For a dataset/benchmark paper, please state release plans and licenses.
  6. [Introduction] The text uses 'Nano Banana Pro' as a synonym for Gemini-3-Pro-Image; define this alias at first use.

Circularity Check

0 steps flagged

No significant circularity: the SOTA claim is evaluated with an external judge and includes an out-of-distribution Reasoning task; remaining concerns are benchmark-provenance risks, not demonstrated circularity.

full rationale

The paper's central derivation is training-data construction -> two-stage fine-tuning -> benchmark evaluation. No equation in the paper defines a predicted quantity in terms of a fitted parameter or benchmark label. The main tables are scored by Gemini-3-Pro, an external VLM, not by the paper's own Qwen3-VL reward models; the RL stage optimizes Qwen3-VL-based rewards (Eqs. 6-8) while evaluation uses Gemini-3-Pro, so the headline IA/TC/BP numbers are not the training objective by construction. The benchmark also includes a Reasoning category explicitly 'beyond the operation types found in the training set', providing at least one out-of-distribution evaluation slice. Self-citations (e.g., refs. [83,84,87] in Sec. 2.2) are related-work pointers to the authors' prior layout/text-rendering papers and are not load-bearing for the claimed result. The legitimate risks noted by a skeptical reader - the benchmark construction in Sec. 4.3 is not described, so train/test overlap from the same HTML/unstructured pipeline cannot be ruled out, and the same Gemini model family is used for data verification and evaluation - are validity/contamination concerns, not circular reductions; the paper provides no statement that test cases are derived from the training pipeline. Under the rule requiring a quoted equation or fitted-parameter reduction, no circular step is identifiable.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on the validity of the proposed benchmark and the unverified VLM judge, plus standard RL/diffusion machinery. No new physical or conceptual entities are introduced; the glyph image is a rendered conditioning signal, not a new theoretical object.

free parameters (3)
  • reward weights λ_acc, λ_cla, λ_pre, λ_qua = not reported
    Weights in Eq. 7 that determine the composite reward in RL; chosen by hand, not learned or reported, but they affect the trained policy.
  • RL guidance coefficient β = not reported
    Hyperparameter in Eqs. 4-5 controlling the contrastive strength between positive and negative branches.
  • candidate count K = not reported
    Number of samples drawn per conditioning input in RL (Eq. 3 context); affects variance of the normalization in Eq. 8.
axioms (4)
  • domain assumption VLM-based scoring (Gemini-3-Pro) is a valid proxy for human judgment of editing quality
    The entire benchmark evaluation rests on this; no correlation to human ratings is shown.
  • domain assumption HTML conversion and rendering preserve source image content faithfully for structured data
    The pixel-perfect non-target claim depends on the VLM's image-to-HTML conversion accuracy, which is not verified.
  • domain assumption The editing model (Qwen-Image-Edit-2509) is a suitable base for text-centric editing
    The method builds on this base; the paper assumes its architecture and pretraining are the right foundation.
  • standard math Standard flow-matching / diffusion background (rectified flow, LoRA, DiffusionNFT)
    Equations 1-5 and the RL objective are taken from the cited literature without derivation.

pith-pipeline@v1.3.0-alltime-deepseek · 30992 in / 11708 out tokens · 101163 ms · 2026-08-02T18:20:14.352186+00:00 · methodology

0 comments
read the original abstract

Instruction-based image editing aims to modify specific content within existing images according to user-provided instructions while preserving non-target regions. Beyond traditional object- and style-centric manipulation, text-centric image editing focuses on modifying, translating, or rearranging textual elements embedded within images. However, existing leading models often struggle to execute complex text editing precisely, frequently producing blurry or hallucinated characters. We attribute these failures primarily to the lack of specialized training paradigms tailored for text-centric editing, as well as the absence of large-scale datasets and standardized benchmarks necessary for a closed-loop training and evaluation system. To address these limitations, we present WeEdit, a systematic solution encompassing a scalable data construction pipeline, two benchmarks, and a tailored two-stage training strategy. Specifically, we propose a novel HTML-based automatic editing pipeline, which generates 330K training pairs covering diverse editing operations and 15 languages, accompanied by standardized bilingual and multilingual benchmarks for comprehensive evaluation. On the algorithmic side, we employ glyph-guided supervised fine-tuning to inject explicit spatial and content priors, followed by a multi-objective reinforcement learning stage to align generation with instruction adherence, text clarity, and background preservation. Extensive experiments demonstrate that WeEdit outperforms previous open-source models by a clear margin across diverse editing operations.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

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

  1. GMO-E$^2$DIT: Grounded Multi-Operation Editing for E-Commerce Images

    cs.CV 2026-07 unverdicted novelty 6.0

    GMO-E²DIT is an agentic framework that decouples VLM-based edit planning from mask-conditioned rendering using reflection loops for reliable multi-operation e-commerce image editing.

  2. GMO-E$^2$DIT: Grounded Multi-Operation Editing for E-Commerce Images

    cs.CV 2026-07 unverdicted novelty 6.0

    GMO-E²DIT is an agentic editing framework that decouples VLM-based planning from mask-conditioned rendering and uses reflection to execute multi-operation e-commerce image edits with error recovery.

  3. TextSculptor: Training and Benchmarking Scene Text Editing

    cs.CV 2026-05 unverdicted novelty 6.0

    TextSculptor supplies an automated data synthesis pipeline yielding 3.2M samples plus a four-task benchmark that raises open-source scene text editing performance.

Reference graph

Works this paper leans on

95 extracted references · 32 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Arial.https://learn.microsoft.com/en-us/typography/fontlist/arial

  2. [2]

    Beautifulsoup4.https://pypi.org/project/beautifulsoup4/

  3. [3]

    Pillow.https://github.com/python-pillow/Pillow

  4. [4]

    Playwright.https://github.com/microsoft/playwright

  5. [5]

    Tailwind.https://tailwindcss.com/

  6. [6]

    Seedream4.5.https://seed.bytedance.com/en/seedream4_5, 2025

    2025

  7. [7]

    Gemini-2.5-flash-image.https://developers.googleblog.com/introducing-gemini-2-5-flash-image/, 2025

    2025

  8. [8]

    Gpt-image-1.5.https://developers.openai.com/api/docs/models/gpt-image-1.5, 2025

    2025

  9. [9]

    Qwen-image-edit-2509.https://huggingface.co/Qwen/Qwen-Image-Edit-2509, 2025

    2025

  10. [10]

    Qwen-image-edit-2511.https://huggingface.co/Qwen/Qwen-Image-Edit-2511, 2025

    2025

  11. [11]

    Gemini-3.0-pro.https://blog.google/products-and-platforms/products/gemini/gemini-3/, 2025.11

    2025

  12. [12]

    https://blog.google/innovation-and-ai/technology/developers-tools/ gemini-3-pro-image-developers/, 2025.11

    Gemini-3.0-pro-image. https://blog.google/innovation-and-ai/technology/developers-tools/ gemini-3-pro-image-developers/, 2025.11

    2025

  13. [13]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023

    2023

  15. [15]

    Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer

    Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer. arXiv preprint arXiv:2505.22705, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Hunyuanimage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, Ying Dong, Kipper Gong, Tianpeng Gu, Xiusen Gu, et al. Hunyuanimage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Bytemorph: Benchmarking instruction-guided image editing with non-rigid motions

    Di Chang, Mingdeng Cao, Yichun Shi, Bo Liu, Shengqu Cai, Shijie Zhou, Weilin Huang, Gordon Wetzstein, Mohammad Soleymani, and Peng Wang. Bytemorph: Benchmarking instruction-guided image editing with non-rigid motions. arXiv preprint arXiv:2506.03107, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Textdiffuser: Diffusion models as text painters

    Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. NeurIPS, 2023

    2023

  19. [19]

    Textdiffuser-2: Unleashing the power of language models for text rendering

    Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. InECCV, 2024

    2024

  20. [20]

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

    Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024

    2024

  21. [21]

    Sharegpt-4o-image: Aligning multimodal models with gpt-4o-level image generation

    Junying Chen, Zhenyang Cai, Pengcheng Chen, Shunian Chen, Ke Ji, Xidong Wang, Yunjin Yang, and Benyou Wang. Sharegpt-4o-image: Aligning multimodal models with gpt-4o-level image generation. arXiv preprint arXiv:2506.18095, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Yufeng Cui, Honghao Chen, Haoge Deng, Xu Huang, Xinghang Li, Jirong Liu, Yang Liu, Zhuoyan Luo, Jin- sheng Wang, Wenxuan Wang, et al. Emu3. 5: Native multimodal models are world learners.arXiv preprint arXiv:2510.26583, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

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

    Pith/arXiv arXiv 2025

  24. [24]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. InNeurIPS, 2021. 22

    2021

  25. [25]

    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 ICML, 2024

    2024

  26. [26]

    Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers

    Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Seedream 3.0 technical report.arXiv preprint arXiv:2504.11346, 2025

    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

    Pith/arXiv arXiv 2025

  28. [28]

    Unireditbench: A unified reasoning-based image editing benchmark.arXiv preprint arXiv:2511.01295, 2025

    Feng Han, Yibin Wang, Chenglin Li, Zheming Liang, Dianyi Wang, Yang Jiao, Zhipeng Wei, Chao Gong, Cheng Jin, Jingjing Chen, et al. Unireditbench: A unified reasoning-based image editing benchmark.arXiv preprint arXiv:2511.01295, 2025

    arXiv 2025

  29. [29]

    Prompt-to-prompt image editing with cross-attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. InICLR, 2023

    2023

  30. [30]

    Denoising diffusion probabilistic models

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

    2020

  31. [31]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InICLR, 2022

    2022

  32. [32]

    In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers.arXiv preprint arXiv:2410.23775, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Hq-edit: A high-quality dataset for instruction-based image editing

    Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Cihang Xie, and Yuyin Zhou. Hq-edit: A high-quality dataset for instruction-based image editing. InICLR, 2024

    2024

  34. [34]

    Leopard: A vision language model for text-rich multi-image tasks

    Mengzhao Jia, Wenhao Yu, Kaixin Ma, Tianqing Fang, Zhihan Zhang, Siru Ouyang, Hongming Zhang, Dong Yu, and Meng Jiang. Leopard: A vision language model for text-rich multi-image tasks. arXiv preprint arXiv:2410.01744, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  36. [36]

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

    2023

  37. [37]

    Flux.1-dev.https://blackforestlabs.ai/announcing-black-forest-labs, 2024

    Black Forest Labs. Flux.1-dev.https://blackforestlabs.ai/announcing-black-forest-labs, 2024

    2024

  38. [38]

    Flux.2-dev.https://huggingface.co/black-forest-labs/FLUX.2-dev, 2025

    Black Forest Labs. Flux.2-dev.https://huggingface.co/black-forest-labs/FLUX.2-dev, 2025

    2025

  39. [39]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    What matters when building vision-language models? NeurIPS, 2024

    Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? NeurIPS, 2024

    2024

  41. [41]

    Unlocking the conversion of web screenshots into html code with the websight dataset.arXiv preprint arXiv:2403.09029, 2024

    Hugo Laurençon, Léo Tronchon, and Victor Sanh. Unlocking the conversion of web screenshots into html code with the websight dataset.arXiv preprint arXiv:2403.09029, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Mixgrpo: Unlocking flow-based grpo efficiency with mixed ode-sde.arXiv preprint arXiv:2507.21802, 2025

    Junzhe Li, Yutao Cui, Tao Huang, Yinping Ma, Chun Fan, Miles Yang, and Zhao Zhong. Mixgrpo: Unlocking flow-based grpo efficiency with mixed ode-sde.arXiv preprint arXiv:2507.21802, 2025

    Pith/arXiv arXiv 2025

  43. [43]

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

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding.arXiv e-prints, 2024

    2024

  44. [44]

    Uniworld-v2: Reinforce image editing with diffusion negative-aware finetuning and mllm implicit feedback

    Zongjian Li, Zheyuan Liu, Qihui Zhang, Bin Lin, Feize Wu, Shenghai Yuan, Zhiyuan Yan, Yang Ye, Wangbo Yu, Yuwei Niu, et al. Uniworld-v2: Reinforce image editing with diffusion negative-aware finetuning and mllm implicit feedback. arXiv preprint arXiv:2510.16888, 2025

    Pith/arXiv arXiv 2025

  45. [45]

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

    Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, et al. Uniworld-v1: High-resolution semantic encoders for unified visual understanding and generation. arXiv preprint arXiv:2506.03147, 2025. 23

    Pith/arXiv arXiv 2025

  46. [46]

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

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

    Pith/arXiv arXiv 2024

  47. [47]

    Flow-grpo: Training flow matching models via online rl

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di ZHANG, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  48. [48]

    Step1x-edit: A practical framework for general image editing.arXiv preprintarXiv:2504.17761, 2025

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprintarXiv:2504.17761, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InICLR, 2023

    2023

  50. [50]

    Glyph-byt5: A customized text encoder for accurate visual text rendering

    Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. InECCV, 2024

    2024

  51. [51]

    Glyph-byt5- v2: A strong aesthetic baseline for accurate multilingual visual text rendering.arXiv preprint arXiv:2406.10208, 2024

    Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Lin Liang, Lijuan Wang, Ji Li, and Yuhui Yuan. Glyph-byt5- v2: A strong aesthetic baseline for accurate multilingual visual text rendering.arXiv preprint arXiv:2406.10208, 2024

    Pith/arXiv arXiv 2024

  52. [52]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.NeurIPS, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.NeurIPS, 2022

    2022

  53. [53]

    Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation.arXiv preprint arXiv:2303.17870, 2023

    Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation.arXiv preprint arXiv:2303.17870, 2023

    Pith/arXiv arXiv 2023

  54. [54]

    X2edit: Revisiting arbitrary- instruction image editing through self-constructed data and task-aware representation learning.arXiv preprint arXiv:2508.07607, 2025

    Jian Ma, Xujie Zhu, Zihao Pan, Qirong Peng, Xu Guo, Chen Chen, and Haonan Lu. X2edit: Revisiting arbitrary- instruction image editing through self-constructed data and task-aware representation learning.arXiv preprint arXiv:2508.07607, 2025

    arXiv 2025

  55. [55]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InCVPR, 2023

    2023

  56. [56]

    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. InICLR, 2024

    2024

  57. [57]

    Pico-banana-400k: A large-scale dataset for text-guided image editing.arXiv preprint arXiv:2510.19808, 2025

    Yusu Qian, Eli Bocek-Rivele, Liangchen Song, Jialing Tong, Yinfei Yang, Jiasen Lu, Wenze Hu, and Zhe Gan. Pico-banana-400k: A large-scale dataset for text-guided image editing.arXiv preprint arXiv:2510.19808, 2025

    arXiv 2025

  58. [58]

    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. InCVPR, 2022

    2022

  59. [59]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InCVPR, 2023

    2023

  60. [60]

    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. InNeurIPS, 2022

    2022

  61. [61]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InICLR, 2021

    2021

  62. [62]

    MIT press, 2018

    Richard S Sutton and Andrew G Barto.Reinforcement learning: An introduction. MIT press, 2018

    2018

  63. [63]

    Ominicontrol: Minimal and universal control for diffusion transformer

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. InICCV, pages 14940–14950, 2025

    2025

  64. [64]

    Longcat-image technical report.arXiv preprint arXiv:2512.07584, 2025

    Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, et al. Longcat-image technical report.arXiv preprint arXiv:2512.07584, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Firered-image-edit-1.0 techinical report.arXiv preprint arXiv:2602.13344, 2026

    Super Intelligence Team, Changhao Qiao, Chao Hui, Chen Li, Cunzheng Wang, Dejia Song, Jiale Zhang, Jing Li, Qiang Xiang, Runqi Wang, et al. Firered-image-edit-1.0 techinical report.arXiv preprint arXiv:2602.13344, 2026

    arXiv 2026

  66. [66]

    Anytext: Multilingual visual text generation and editing.arXiv preprint arXiv:2311.03054, 2023

    Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing.arXiv preprint arXiv:2311.03054, 2023. 24

    Pith/arXiv arXiv 2023

  67. [67]

    I2i-bench: A comprehensive benchmark suite for image-to-image editing models.arXiv preprint arXiv:2512.04660, 2025

    Juntong Wang, Jiarui Wang, Huiyu Duan, Jiaxiang Kang, Guangtao Zhai, and Xiongkuo Min. I2i-bench: A comprehensive benchmark suite for image-to-image editing models.arXiv preprint arXiv:2512.04660, 2025

    arXiv 2025

  68. [68]

    Gpt-image-edit-1.5 m: A million-scale, gpt-generated image dataset.arXiv preprint arXiv:2507.21033, 2025

    Yuhan Wang, Siwei Yang, Bingchen Zhao, Letian Zhang, Qing Liu, Yuyin Zhou, and Cihang Xie. Gpt-image-edit-1.5 m: A million-scale, gpt-generated image dataset.arXiv preprint arXiv:2507.21033, 2025

    Pith/arXiv arXiv 2025

  69. [69]

    Omniedit: Building image editing generalist models through specialist supervision

    Cong Wei, Zheyang Xiong, Weiming Ren, Xeron Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image editing generalist models through specialist supervision. InICLR, 2024

    2024

  70. [70]

    Chain-of-thought prompting elicits reasoning in large language models.NeurIPS, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.NeurIPS, 2022

    2022

  71. [71]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  73. [73]

    Chronoedit: Towards temporal reasoning for image editing and world simulation

    Jay Zhangjie Wu, Xuanchi Ren, Tianchang Shen, Tianshi Cao, Kai He, Yifan Lu, Ruiyuan Gao, Enze Xie, Shiyi Lan, Jose M Alvarez, et al. Chronoedit: Towards temporal reasoning for image editing and world simulation. arXiv preprint arXiv:2510.04290, 2025

    arXiv 2025

  74. [74]

    Editreward: A human-aligned reward model for instruction-guided image editing.arXiv preprint arXiv:2509.26346, 2025

    Keming Wu, Sicong Jiang, Max Ku, Ping Nie, Minghao Liu, and Wenhu Chen. Editreward: A human-aligned reward model for instruction-guided image editing.arXiv preprint arXiv:2509.26346, 2025

    arXiv 2025

  75. [75]

    Less-to-more generalization: Unlocking more controllability by in-context generation

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more generalization: Unlocking more controllability by in-context generation. InICCV, 2025

    2025

  76. [76]

    Omnigen: Unified image generation

    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. InCVPR, 2025

    2025

  77. [77]

    Imagereward: Learning and evaluating human preferences for text-to-image generation.NeurIPS, 2023

    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.NeurIPS, 2023

    2023

  78. [78]

    Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025

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

    Pith/arXiv arXiv 2025

  79. [79]

    Glyphcontrol: Glyph conditional control for visual text generation.NeurIPS, 2023

    Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation.NeurIPS, 2023

    2023

  80. [80]

    Echo-4o: Harnessing the power of gpt-4o synthetic images for improved image generation

    Junyan Ye, Dongzhi Jiang, Zihao Wang, Leqi Zhu, Zhenghao Hu, Zilong Huang, Jun He, Zhiyuan Yan, Jinghua Yu, Hongsheng Li, et al. Echo-4o: Harnessing the power of gpt-4o synthetic images for improved image generation. arXiv preprint arXiv:2508.09987, 2025

    Pith/arXiv arXiv 2025

Showing first 80 references.