Pith. sign in

REVIEW 4 major objections 7 minor 3 cited by

DreamPainter: Image Background Inpainting for E-commerce Scenarios

T0 review · 4 major / 7 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 DreamPainter repaints e-commerce product backgrounds from text plus one reference image, beating five open-source baselines in human preference tests.

desk verdict Solid engineering with a new dataset and sensible two-stage training, but the evaluation's likely train/test distribution overlap keeps the headline gains unproven. read the letter →

arxiv 2508.02155 v1 pith:6YBIRUA2 submitted 2025-08-04 cs.CV cs.AI

Sijie Zhao , Jing Cheng , Yaoyao Wu , Hao Xu , Shaohui Jiao This is my paper
classification cs.CVcs.AI
keywords DreamPaintere-commercebackgroundinpaintingdiffusiontransformerreferenceimageconditioningEcom-400Ktwo-stagetrainingLoRAscalingtext-guided
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

DreamPainter is a diffusion-transformer inpainting system that repaints the background of a product photo while leaving the foreground product unchanged. The paper argues that generic inpainting models fail in e-commerce because they lack domain-specific training data and depend on text alone, so it contributes a 400,000-image dataset and adds an optional reference image as a second control signal. On its own evaluation set, the model reports the highest scores on automated preference metrics (IR-Score 0.841 with text only and 0.932 with a reference image) and wins 88.8%–98.8% of pairwise human comparisons against five open-source baselines. The practical stake is that a designer can generate an appropriate background either from a text description alone or from a text description plus a style example, with a dial controlling how strongly the reference influences the result.

What carries the argument

The load-bearing mechanism is feeding the model a token sequence formed by concatenating the text embedding, the noisy target latent, and the optional reference latent along the spatial dimension, with the reference tokens' rotary positional indices shifted by the target canvas height and width, $(h+i, w+j)$, so the model does not confuse the two images and keeps their 2D arrangement. The auxiliary control branch carries mask and foreground latents into each DiT block through zero-initialized gates, and the two-stage recipe first trains only that branch, then adds rank-256 LoRA to all linear layers while freezing the branch, with 50% reference dropout. At inference, LoRA scaling and attention-score modulation provide a continuous similarity knob for how strongly the reference image guides the output.

What would settle it

Take a held-out set of product photos scraped from real online stores, with manually checked masks and independently photographed reference images that were not produced by the same text-to-image and editing pipeline, and rerun the T2I and TR2I comparisons; if the IR-Score and human win-rate gaps over the baselines shrink to near zero, the claimed superiority is an artifact of the shared synthetic distribution.

Watch

Extended reading notes

Core claim

The central claim is that a text-to-image diffusion transformer can be turned into an e-commerce background-inpainting model by adding two condition channels without sacrificing either. The foreground product and mask are injected through a zero-initialized auxiliary control branch into the DiT layers; the reference image latent is concatenated along the spatial dimension with the noisy target latent and the text embedding, and is given shifted rotary position indices $(h+i, w+j)$ so the model treats the reference as a canvas placed beside the target. Training happens in two stages: first text-only control is learned with the control branch, then the branch is frozen and LoRA modules adapt the DiT blocks to reference-conditioned inputs, with reference information randomly dropped at 50% to keep the text-only capability. At inference, the user can scale the LoRA output or modulate attention scores to balance reference similarity against generative randomness. The paper reports that this setup substantially outperforms current open-source inpainting and multi-image editing models on both tasks while preserving product identity, with object consistency of 0.986 in T2I and 0.992 in TR2I.

Load-bearing premise

The entire performance comparison presumes that the evaluation images are representative real e-commerce product photos rather than samples that share the same synthetic generation style as the training images.

Editorial extensions

If this is right

  • A single deployed model can serve both text-only and text-plus-reference background generation, because the 50% reference dropout in stage two preserves the text-only path.
  • Product fidelity stays high in both modes (object consistency 0.986 in T2I and 0.992 in TR2I), so the foreground is not redrawn, duplicated, or distorted when the background changes.
  • Users get a continuous similarity dial through LoRA scaling and attention-score modulation, making it possible to reuse a reference style without copying it exactly.
  • The cost of the approach sits in training on the synthetic DreamEcom-400K pipeline; at inference the model needs only a product image, a mask, a prompt, and optionally one reference image.

Reading between the lines

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

  • If the evaluation set were replaced by real seller photos that were not generated by the same text-to-image and editing pipeline, the reported gaps (IR-Score 0.841 vs. 0.714 in T2I) would be the direct thing to re-measure; the paper leaves this transfer question open.
  • The shifted-position concatenation trick is a generic way to add visual reference conditioning to a diffusion transformer, so the recipe could plausibly transfer to other DiT-based backbones without dataset-specific changes.
  • Because the training reference images are themselves produced by an editing model, the reference-conditioned path inherits that model's artifacts; human-curated or photographed references would test whether the gains survive independent reference sources.
  • The out-of-domain examples in the appendix cover humans, animals, and vehicles but are shown qualitatively only; adding object-consistency scores for those categories would test whether the generalization claim holds beyond products.
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

4 major / 7 minor

Summary. DreamPainter proposes an e-commerce background inpainting framework built on CogView4, trained on a newly constructed 400K-image dataset (DreamEcom-400K) whose samples contain product images, instance masks, bilingual prompts, and background references generated by Seedream3, GroundingDINO/SAM, and BAGLE. The method introduces an auxiliary control branch for mask/foreground conditioning and a second-stage LoRA injection for reference-image guidance, with two user-adjustable scaling mechanisms. Experiments compare against open-source baselines on 164 T2I and 122 TR2I samples; DreamPainter reports large gains in IR-Score, PickScore, Object Consistency, and human-preference win rates. The main risk is that the evaluation set's provenance is not specified and may be drawn from the same synthetic pipeline used to build the training data.

Significance. If the performance gains transfer to independently sourced e-commerce images, this is a solid applied contribution: the dataset fills an identified data gap, the two-stage training protocol is sensible, and the reference-control knobs are genuine inference-time user controls rather than fitted parameters. I found no equation-level circularity. The paper is weakened by an underspecified and likely non-independent evaluation set, a malformed object-consistency formula, and lack of variance/significance reporting; these issues make the current evidence for 'significantly outperforms' incomplete.

major comments (4)
  1. [§5.1, Evaluation Details (and §3)] The evaluation set is described only as 'based on real-world e-commerce scenarios and product images,' with no information about how the 164 T2I and 122 TR2I product images, prompts, masks, and reference images were created. Since the training set in §3 is generated by Seedream3, BAGLE, and GroundingDINO/SAM, the central performance claim is not yet interpretable: if the evaluation samples come from the same pipeline, DreamPainter has been tested on its training distribution while baselines have not. Please specify the evaluation-set construction pipeline, state whether it overlaps with DreamEcom-400K, and verify the main comparisons on an independent, realistic set of product photos (e.g., real catalog images with human-annotated masks and references).
  2. [§5.1, Object Consistency metric] The formula printed as '1−P(Mgen /∈Mgt)P(Mgt)' is malformed: the events are not defined (Mgen, Mgt), the expression is not a standard probability or overlap score, and the values 0.986 and 0.992 in Tables 4a and 6a cannot be reproduced from it. This metric is load-bearing for the product-consistency claim, so please define the masks, give a correct set-theoretic or probabilistic expression, and describe how the segmentation masks are compared; include any intersection/union or distance computation used.
  3. [§5.2–5.3, Tables 4a and 6a] All quantitative results are reported as point estimates without standard deviations, confidence intervals, or significance tests. With only 164 and 122 evaluation samples, the large margins (e.g., IR-Score 0.932 vs 0.393 in TR2I) could still be accompanied by wide intervals; please report per-sample distributions, bootstrap confidence intervals, or paired significance tests for the main metrics.
  4. [§5.2–5.3, Human Evaluation] The human study reports only aggregate win rates from four evaluators, with no information about the number of pairwise comparisons per condition, the instructions given, or inter-rater agreement. Given the small rater pool and small evaluation set, win rates of 88.8–98.8% need confidence intervals or per-evaluator breakdowns before the claim of significant human preference is supported.
minor comments (7)
  1. [Abstract] The abstract contains the typo 'genenation'; it should be 'generation'.
  2. [§4.1, Eq. (2)] In Eq. (2), the control branch input lists Z^t_tgt twice; please clarify whether one of these should be a different latent (e.g., Z_tgt).
  3. [§5.2] The baseline list numbers two entries as '(4)'; renumber the list to avoid ambiguity.
  4. [Table 4a] Ours has a lower CLIP-T score (0.337) than UNO (0.345); the text's summary only highlights IR-Score and PickScore, but readers should be told why the method does not lead on CLIP-T.
  5. [§5.4] The conclusion that attention modulation preserves quality at low scales while LoRA scaling degrades it is based only on the qualitative grid of Fig. 8; a quantitative ablation would support the claim.
  6. [Appendix A.1] The claim of 'robust performance' on out-of-domain subjects is asserted from unquantified examples; please add quantitative generalization results or soften the claim.
  7. [Reproducibility] No code, model weights, or evaluation-set release is mentioned; for reproducibility and for the dataset to serve as a benchmark, please state a release plan.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; one minor non-load-bearing self-citation only.

full rationale

DreamPainter's central claims are empirical: the model is trained on its own DreamEcom-400K dataset and evaluated against external open-source baselines using external metrics (IR-Score, PickScore, CLIP) plus human raters. The reference-influence knobs (LoRA scale and attention scale) are user-controlled parameters, not values fitted to the evaluation data, so no prediction reduces to a fitted input. The architecture choices (control branch, spatial concatenation of reference tokens, two-stage training) are presented as design decisions with citations to prior work, not as derived predictions. No equation in Section 4 is self-referential, and the malformed Object Consistency formula in Section 5.1 is a typographical issue, not a circular construction. The only author-overlap citation is [7] (Seed-X), which includes co-author Sijie Zhao; it is cited in the Related Work discussion of unified multimodal frameworks and is not load-bearing for any experimental claim. The reader-flagged concern that the evaluation set may share the synthetic Seedream3/BAGLE/GroundingDINO pipeline with the training set is not evidenced in the paper: Section 5.1 states the evaluation set is 'based on real-world e-commerce scenarios and product images,' and no construction detail indicates that the 164 T2I or 122 TR2I evaluation samples were generated by the authors' own pipeline. Without such evidence, distribution overlap is a correctness-risk hypothesis rather than a demonstrated circular step. Accordingly, the derivation chain is self-contained against external benchmarks, aside from one minor, non-load-bearing self-citation.

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

The central claims rest on a synthetic data pipeline whose components (Seedream3, BAGLE) are not independently verifiable from the paper, and on an evaluation protocol that may share the training distribution.

free parameters (3)
  • LoRA rank = 256
    Architecture hyperparameter chosen in Sec 5.1; the reference-image integration mechanism depends on this capacity.
  • Reference sampling rate = 50%
    Stage 2 training uses reference images at a 50% sampling rate; this controls the balance between text-only and text-plus-reference capability.
  • Inference scale s = user-adjustable
    The similarity-to-reference dial in Eqs. 5-7; not learned, but central to the flexible control claim.
assumptions (3)
  • domain assumption Seedream3 generates high-quality, artifact-free product images suitable as supervised training targets.
    Sec 3 Step 2 states that Seedream3 images need no post-processing; the quality of the entire dataset rests on this unverified assumption.
  • domain assumption BAGLE editing removes products and their shadows/reflections with high fidelity, producing clean background reference images.
    Sec 3 Step 4; reference image quality is central to TR2I training and evaluation.
  • domain assumption The automated metrics (IR-Score, PickScore, CLIP score, Object Consistency) reflect human judgment in this domain.
    Sec 5.1 uses these metrics without validation on e-commerce data; the Object Consistency formula is additionally malformed as printed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DreamPainter: Image Background Inpainting for E-commerce Scenarios." pith.science (2026). https://pith.science/paper/6YBIRUA2

@misc{pith2026250802155,
  author       = {Pith},
  title        = {Pith review of: DreamPainter: Image Background Inpainting for E-commerce Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YBIRUA2}},
  note         = {Machine review of arXiv:2508.02155}
}
read the original abstract

Although diffusion-based image genenation has been widely explored and applied, background generation tasks in e-commerce scenarios still face significant challenges. The first challenge is to ensure that the generated products are consistent with the given product inputs while maintaining a reasonable spatial arrangement, harmonious shadows, and reflections between foreground products and backgrounds. Existing inpainting methods fail to address this due to the lack of domain-specific data. The second challenge involves the limitation of relying solely on text prompts for image control, as effective integrating visual information to achieve precise control in inpainting tasks remains underexplored. To address these challenges, we introduce DreamEcom-400K, a high-quality e-commerce dataset containing accurate product instance masks, background reference images, text prompts, and aesthetically pleasing product images. Based on this dataset, we propose DreamPainter, a novel framework that not only utilizes text prompts for control but also flexibly incorporates reference image information as an additional control signal. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, maintaining high product consistency while effectively integrating both text prompt and reference image information.

Figures

Figures reproduced from arXiv: 2508.02155 by the authors.

Figure 1
Figure 1. Overview of DreamPainter, which can not only perform background inpainting based on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Data samples from DreamEcom-400k (left) and construction pipeline (right). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Framework of DreamPainter. We inject mask and foreground information through the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of T2I Inpainting between different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of TR2I Inpainting between different methods. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Qualitative analyses between Attention modulation and LoRA scaling. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results of T2I background inpainting on out-of-domain tasks. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Design Your Ad: Personalized Advertising Image and Text Generation with Unified Autoregressive Models

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    Uni-AdGen uses a unified autoregressive framework with foreground perception, instruction tuning, and coarse-to-fine preference modules to generate personalized image-text ads from noisy user behaviors, outperforming ...

  2. Self-Reasoning Agentic Framework for Narrative Product Grid-Collage Generation

    cs.CV 2026-04 unverdicted novelty 5.0 of 10

    A self-reasoning agentic framework constructs a Product Narrative Framework, generates constraint-aware unified grid collages, and refines outputs via failure attribution to improve narrative coherence and aesthetics ...

  3. Pinterest Canvas: Large-Scale Image Generation at Pinterest

    cs.CV 2026-03 conditional novelty 4.0 of 10

    A FLUX-style base diffusion model plus task-specific fine-tunes and product-preserving pipelines yields double-digit Pinterest ads engagement lifts and higher no-defect rates than GPT-Image, FLUX Kontext, and Nano Banana.

Reference graph

Works this paper leans on

36 extracted references · 10 canonical work pages · cited by 3 Pith papers

  1. [1]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 18392–18402, 2023

    work page 2023

  2. [2]

    Ecomxl controlnet inpaint

    Alimama Creative. Ecomxl controlnet inpaint. https://huggingface.co/alimama-creative/ EcomXL_controlnet_inpaint, 2024

    work page 2024

  3. [3]

    Flux.1-dev-controlnet-inpainting

    Alimama Creative. Flux.1-dev-controlnet-inpainting. https://github.com/alimama-creative/ FLUX-Controlnet-Inpainting , 2024

    work page 2024

  4. [4]

    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

    arXiv 2025

  5. [5]

    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 learning, 2024

    2024

  6. [6]

    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

    arXiv 2025

  7. [7]

    Seed-x: Multimodal models with unified multi-granularity comprehension and generation

    Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024

    arXiv 2024

  8. [8]

    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

Show all 36 references
  1. [9]

    Vace: All-in-one video creation and editing

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. Vace: All-in-one video creation and editing. arXiv preprint arXiv:2503.07598, 2025

    2025 arXiv

  2. [10]

    Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion

    Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, pages 150–168. Springer, 2024

    2024

  3. [11]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  4. [12]

    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

  5. [13]

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

    2024

  6. [14]

    Flux fill

    Black Forest Labs. Flux fill. https://bfl.ai/announcements/24-11-21-tools , 2024

    2024

  7. [15]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023

    2023 arXiv

  8. [16]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022

    2022

  9. [17]

    T2i- adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i- adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 4...

    2024

  10. [18]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023

    2023 arXiv

  11. [19]

    Controlnext: Powerful and efficient control for image and video generation

    Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 11

    2024 arXiv

  12. [20]

    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

  13. [21]

    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

  14. [22]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  15. [23]

    High-resolution image synthesis with latent diffusion models, 2021

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021

    2021

  16. [24]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021

    2021 arXiv

  17. [25]

    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. arXiv preprint arXiv:2411.15098, 2024

    2024 arXiv

  18. [26]

    High-fidelity pluralistic image completion with transformers

    Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transformers. arXiv preprint arXiv:2103.14031, 2021

    2021 arXiv

  19. [27]

    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

  20. [28]

    Less-to-more gener- alization: Unlocking more controllability by in-context generation

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more gener- alization: Unlocking more controllability by in-context generation. arXiv preprint arXiv:2504.02160, 2025

    2025 arXiv

  21. [29]

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

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

    2023

  22. [30]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  23. [31]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023

    2023

  24. [32]

    Large scale image completion via co-modulated generative adversarial networks

    Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428, 2021

    2021 arXiv

  25. [33]

    Omdet: Large-scale vision-language multi-dataset pre- training with multimodal detection network

    Tiancheng Zhao, Peng Liu, and Kyusong Lee. Omdet: Large-scale vision-language multi-dataset pre- training with multimodal detection network. IET Computer Vision, 2024

    2024

  26. [34]

    Cogview3: Finer and faster text-to-image generation via relay diffusion

    Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. In European Conference on Computer Vision, pages 1–22. Springer, 2024

    2024

  27. [35]

    Cogview4

    ZhipuAI. Cogview4. https://github.com/THUDM/CogView4, 2025

    2025

  28. [36]

    A task is worth one word: Learning with task prompts for high-quality versatile image inpainting

    Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision, pages 195–211. Springer, 2024. 12 A Appendix A.1 Qualitative Results o...

    2024

Pith tools

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