Pith. sign in

REVIEW 3 major objections 6 minor 48 references

A shared multimodal diffusion base, rapidly specialized per task, produces product-safe image edits that raise engagement by double digits.

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 · grok-4.5

2026-07-15 13:49 UTC pith:UD7Y77AV

load-bearing objection Competent industrial systems paper: shared multimodal base plus specialized fine-tunes and hard compositing deliver real A/B engagement lifts, but third-party no-defect wins are inflated by asymmetric product locks. the 3 major comments →

arxiv 2603.06453 v2 pith:UD7Y77AV submitted 2026-03-06 cs.CV

Pinterest Canvas: Large-Scale Image Generation at Pinterest

Yu Wang , Eric Tzeng , Raymond Shiau , Jie Yang , Dmitry Kislyuk , Charles Rosenberg This is my paper
classification cs.CV
keywords Image GenerationDiffusion ModelsMultimodal Understandingoutpaintingproduct preservationtask-specific fine-tuningengagement lift
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.

Generic image generators are hard to control for production jobs that forbid altering real products. Pinterest Canvas instead trains one foundational diffusion model on a large multimodal mix of editing tasks, then fine-tunes dedicated variants on task-specific data so each use case can enforce its own hard constraints. For background enhancement and aspect-ratio outpainting the resulting models strictly preserve the original product via masks and compositing, beat third-party systems on human no-defect rates, and deliver 18.0% and 12.5% click-through lifts in live A/B tests. The same base also supports multi-image scene synthesis and short image-to-motion clips, showing the pattern generalizes. The practical takeaway is that strict product requirements are better met by specialized descendants of a strong shared model than by prompting a single general system.

Core claim

Training one broad multimodal diffusion model once, then rapidly fine-tuning task-specific variants on curated data with mask supervision and original-product compositing, yields controllable editors that satisfy strict product-preservation rules, outperform third-party models on human defect rates, and raise engagement 18.0% CTR for background outpainting and 12.5% CTR for aspect-ratio outpainting.

What carries the argument

The Canvas family: a FLUX.1-Kontext-style multimodal DiT first trained jointly on billions of text+image editing examples (multi-view, outpainting, scene synthesis, etc.), then specialized by fine-tuning with task prefixes, mask-based supervision, and post-generation compositing of the original high-resolution product cutout.

Load-bearing premise

The reported superiority and engagement lifts come mainly from the specialized model weights rather than from the full production stack of eligibility filters, multi-candidate generation, reward ranking, seed tuning, original-product compositing, and dual human review that third-party models did not receive.

What would settle it

Re-run the same 996-product human evaluation after stripping Canvas of product compositing and multi-candidate reward ranking, and after running the third-party models through the full Canvas pipeline (masks, compositing, multi-gen, reward filter); if the no-defect gap vanishes or reverses, the model-centric claim is falsified.

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

If this is right

  • Advertisers can enhance catalog images at scale while guaranteeing the product itself is never altered or extended.
  • Sharing one base model across tasks reduces training cost (17% in the paper’s comparison) without harming downstream no-defect rates.
  • The same base converges quickly on multi-image scene synthesis and short image-to-motion generation.
  • Live CTR, gCTR30, and click volume rise double-digit when Canvas-enhanced images replace originals on Home Feed, Related Pins, and Search.
  • A rater-trained reward model plus multi-candidate generation makes human review cost-effective by raising yield.

Where Pith is reading between the lines

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

  • Platforms that own large, high-quality visual catalogs can systematically out-perform public general models on identity-critical vertical tasks by controlling both pretraining mix and post-processing.
  • The multi-candidate + reward-ranking + dual-rater pipeline may contribute as much to production yield as the fine-tuned weights themselves; pure model-vs-model comparisons understate the real gap.
  • The base-then-specialize pattern should transfer to other domains that demand non-negotiable identity preservation (medical imaging, legal evidence, e-commerce fashion).
  • As more product-relevant tasks are added to the joint base mix, positive transfer is likely to keep accelerating new-variant convergence.

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 Pinterest Canvas, a production image-generation system that first trains a multimodal diffusion base (FLUX.1 Kontext-style DiT with flow matching) on a large mix of text–image and (text+image)–image editing tasks, then rapidly fine-tunes task-specific variants. Case studies focus on ads background outpainting and aspect-ratio outpainting, with pipeline components for product-mask training, post-generation product compositing, an outpainting VAE, multi-candidate generation, rater-trained reward ranking, eligibility filters, and dual human review. Offline dual-rater evaluations (N=996) report higher overall no-defect rates than GPT-Image 1.5, FLUX.1 Kontext, and Nano Banana; online A/B tests report +18.0% CTR / +18.6% click volume for background outpainting and +12.5% CTR / +12.9% click volume for aspect-ratio outpainting. Additional variants (multi-image scene synthesis, image-to-motion) are shown qualitatively.

Significance. If the results hold under fairer comparison, this is a strong industrial systems contribution for KDD: it documents a practical base-then-specialize recipe for generative models under strict product constraints (identity preservation, ads compliance, opt-out retraining), with large-scale online engagement evidence and concrete training/inference practices (timestep shift sweeps, multimodal CFG variants, β2/EMA stability). The shared-base cost reduction (17%) and positive-transfer ablation are useful for practitioners. The work is less a novel generative architecture than a carefully engineered production framework; its value is in the end-to-end control story and measured product impact rather than new theory.

major comments (3)
  1. §5.2 / Table 1: The claim that Canvas “outperforms third-party models” on background outpainting is not a like-for-like model comparison. §4.1 states that after diffusion Canvas composites the original high-resolution product cutout back onto the generated background, and training already uses InSPyReNet masks so the model does not extend the product. Third-party systems receive only the same prompt and must regenerate the full image. Product no-defect rates drive most of the overall gap (Canvas 84.0% vs 52.9–74.6%), while background no-defect rates are essentially tied (Canvas 54.9% vs 50.8–55.9%). The paper notes that candidate ranking was disabled for Canvas, but does not address this compositing asymmetry. Either (i) re-evaluate third parties under an equivalent mask+composite protocol, (ii) report product-free / background-only metrics as primary, or (iii) reframe Table 1 and the ab
  2. §5.1 / §5.3: Attribution of gains to “task-specific fine-tuning of Canvas” vs the full safeguard stack is incomplete. The fine-tuning ablation (54.2% → 58.4% no-defect, N=637) is modest relative to the pipeline (eligibility filters, multi-candidate generation, reward ranking, seed tuning, dual human review, compositing, super-resolution). Online A/B lifts are valid for the production system but do not isolate the generative model. A clearer decomposition—e.g., base vs fine-tuned vs full pipeline yield, and engagement with/without multi-candidate selection—would make the central systems claim load-bearing rather than confounded.
  3. §5.2–5.3 and Appendix D: Statistical reporting is thin for the headline numbers. Table 1 and Table 2 give point percentages without confidence intervals, inter-rater agreement, or significance tests despite large N (996 products; ~16M impressions per arm). Appendix D raw counts allow approximate CTR CIs, but these should appear in the main results so readers can judge whether the 18.0% / 12.5% lifts and the 47.2% vs 42.5% no-defect gap are robust.
minor comments (6)
  1. §3.2.1 Eq. (1): The timeshift formula is useful; state explicitly whether the same √(m/n)=6.30 is used at 1024² or re-tuned, and whether EditReward correlates with the dual-rater product/background rubric used later.
  2. §3.3 Eqs. (4)–(5): Multimodal CFG variants are a practical contribution; a small quantitative table (not only qualitative preference) on the internal 272-pair benchmark would strengthen the “best practices” claim.
  3. Figure 2 and several appendix figures appear heavily corrupted/garbled in the manuscript text (OCR-like blocks). Ensure camera-ready figures are clean and captions self-contained.
  4. §4.1.1 Outpainting VAE: The color-seam fix is important; report a quantitative seam/color metric or rater preference, not only the qualitative Figure 5.
  5. Related work: briefly situate against commercial inpainting/outpainting controlnets and e-commerce-specific systems already cited in the appendix (EcomXL, DreamPainter) in the main Related Work section.
  6. Abstract and §1: “outperform third-party models” should be softened consistently with any Table 1 reframing after revision.

Circularity Check

0 steps flagged

Empirical systems paper with external A/B and human evaluations; no derivation reduces to its own inputs by construction.

full rationale

Pinterest Canvas is an engineering/systems paper: train a multimodal diffusion base, fine-tune task-specific variants, and measure product impact via online A/B CTR/gCTR/click volume and offline human no-defect rates against third-party models. There is no first-principles derivation chain, uniqueness theorem, or fitted parameter that is then re-presented as an independent prediction. The reward models (Appendix C) are trained on historical rater labels and used only for candidate ranking inside the production pipeline; the paper’s load-bearing claims are the dual-rater human reviews and the live A/B engagement lifts, which are external measurements, not quantities forced by those fits. Self-citations (OmniSage, PinCLIP) supply dataset-construction tools, not load-bearing uniqueness results. Asymmetric compositing/masking that favors Canvas on product preservation is a fairness/correctness concern, not circularity. No step reduces Eq. X to Eq. Y by definition or renames a fit as a prediction. Score 0 is the honest finding.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The paper rests on standard diffusion/flow-matching practice plus many operational choices (dataset mix ratios, timeshift, CFG scales, seeds, reward models, human rubrics) that are fitted or hand-chosen on internal data. No new physical entities; the main invented constructs are the Canvas system family and task-specific pipeline components. Claims about superiority and engagement depend on those choices and on proprietary labels more than on parameter-free theory.

free parameters (6)
  • timestep shift √(m/n) at 512² = 6.30
    Swept over {1.65, 1.88, 3.16, 6.30, 12.57} and selected 6.30 by average EditReward on 272 internal pairs (§3.2.1).
  • multimodal CFG scale and variant = s∈{1,3,7}; variants (4)/(5)
    Task-dependent choice of Eq. (4) vs (5) and scales s=7 (editing/background), s=3 (aspect-ratio), s=1 (super-res) selected qualitatively/empirically (§3.3).
  • AdamW β2 = 0.95
    Lowered to 0.95 for training stability after loss-spike experiments (§3.2.2).
  • multimodal dataset task mix ratios = listed mix in §3.1.3
    Hand-tuned percentages (e.g., multi-view 27%, OmniSage 22.2%, background outpainting 11.7%, …) to balance joint learning (§3.1.3).
  • diffusion seed ranking for production = top-ranked subset of 1024
    Offline sweep of 1,024 seeds on 100 inputs ranked by average reward; top seeds fixed in production (§4.2.5).
  • outpainting reward models = fitted on historical rater annotations
    ViT reward heads trained on retrospective dual-rater defect labels; absolute scores used for ranking and post-generation filtering (Appendix C).
axioms (5)
  • domain assumption Flow-matching multimodal DiT (FLUX.1 Kontext-style) with VAE-latent concatenation is an adequate backbone for joint text+image editing.
    Adopted wholesale from cited FLUX/Kontext literature; no independent derivation (§3.2).
  • ad hoc to paper Compositing the original segmented product cutout after generation guarantees product identity for ads compliance.
    Core production safeguard in §4.1; assumes segmentation is accurate and that identity equals pixel paste of the cutout.
  • domain assumption Dual human raters using the Appendix B rubrics define ground-truth defects for both model selection and third-party comparison.
    All offline yield claims and reward training rest on these labels (§4.2.3, Table 1).
  • domain assumption Aggressive filtering of Pinterest corpus (opt-outs, quality, relevance, VLM captions) yields training data that generalizes to production ads imagery.
    Dataset section §3.1; not externally validated.
  • ad hoc to paper Positive transfer from multi-task base pretraining enables rapid specialization without harming outpainting quality.
    Supported only by internal no-defect comparison of multi-task base vs outpainting-only prior version (§5.1).
invented entities (3)
  • Pinterest Canvas base + task-specific variant family no independent evidence
    purpose: Organize shared pretraining and specialized fine-tunes for product image editing under strict requirements.
    System name and design pattern introduced by the paper; evidence is internal A/B and rater studies only.
  • Outpainting VAE decoder (mask + masked image conditioned) no independent evidence
    purpose: Reduce color seams between original content and generated outpaint bands.
    Fine-tuned from FLUX.1 VAE on Pinterest data; shown only with qualitative before/after (§4.1.1).
  • Task-specific outpainting reward models no independent evidence
    purpose: Rank multi-generation candidates and filter likely human-review failures.
    Trained on proprietary rater history; no external benchmark release (Appendix C).

pith-pipeline@v1.1.0-grok45 · 23798 in / 4066 out tokens · 46698 ms · 2026-07-15T13:49:28.623266+00:00 · methodology

0 comments
read the original abstract

While recent image generation models demonstrate a remarkable ability to handle a wide variety of image generation tasks, this flexibility makes them hard to control via prompting or simple inference adaptation alone, rendering them unsuitable for use cases with strict product requirements. In this paper, we introduce Pinterest Canvas, our large-scale image generation system built to support image editing and enhancement use cases at Pinterest. Canvas is first trained on a diverse, multimodal dataset to produce a foundational diffusion model with broad image-editing capabilities. However, rather than relying on one generic model to handle every downstream task, we instead rapidly fine-tune variants of this base model on task-specific datasets, producing specialized models for individual use cases. We describe key components of Canvas and summarize our best practices for dataset curation, training, and inference. We also showcase task-specific variants through case studies on background enhancement and aspect-ratio outpainting, highlighting how we tackle their specific product requirements. Online A/B experiments demonstrate that our enhanced images receive a significant 18.0% and 12.5% engagement lift, respectively, and comparisons with human raters further validate that our models outperform third-party models on these tasks. Finally, we showcase other Canvas variants, including multi-image scene synthesis and image-to-video generation, demonstrating that our approach can generalize to a wide variety of potential downstream tasks.

discussion (0)

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

Reference graph

Works this paper leans on

48 extracted references · 1 canonical work pages

  1. [1]

    Alimama Creative. 2024. EcomXL ControlNet Inpaint. https://huggingface.co/ alimama-creative/EcomXL_controlnet_inpaint

    2024

  2. [2]

    Alimama Creative. 2024. FLUX ControlNet Inpainting. https://github.com/ alimama-creative/FLUX-Controlnet-Inpainting

    2024

  3. [3]

    Anirudhan Badrinath, Alex Yang, Kousik Rajesh, Prabhat Agarwal, Jaewon Yang, Haoyu Chen, Jiajing Xu, and Charles Rosenberg. 2025. OmniSage: Large Scale, Multi-Entity Heterogeneous Graph Representation Learning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25). ACM, 4261–4272. doi:10.1145/3711896.3737253

    doi:10.1145/3711896.3737253 2025

  4. [4]

    Josh Beal, Eric Kim, Jinfeng Rao, Rex Wu, Dmitry Kislyuk, and Charles Rosenberg

  5. [5]

    InProceedings of The ACM Web Conference 2026 (TheWebConf ’26)

    PinCLIP: Large-scale Foundational Multimodal Representation at Pinter- est. InProceedings of The ACM Web Conference 2026 (TheWebConf ’26). ACM, Dubai, United Arab Emirates, 9 pages. https://www.pinterestcareers.com/media/ eoqd5wcs/pinclip.pdf

    2026

  6. [6]

    Black Forest Labs. 2024. FLUX.1. https://github.com/black-forest-labs/flux

    2024

  7. [7]

    2025.FLUX.2: Analyzing and Enhancing the Latent Space of FLUX – Representation Comparison

    Black Forest Labs. 2025.FLUX.2: Analyzing and Enhancing the Latent Space of FLUX – Representation Comparison. https://bfl.ai/research/representation- comparison

    2025

  8. [8]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. 2023. InstructPix2Pix: Learning to Follow Image Editing Instructions. arXiv:2211.09800 [cs.CV] https: //arxiv.org/abs/2211.09800

    Pith/arXiv arXiv 2023

  9. [9]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. 2024. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. arXiv:2403.03206 [cs.CV] https...

    Pith/arXiv arXiv 2024

  10. [10]

    Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Jiaming Wang Cao Li, Zengqi Xun, Chengyue Sun, Rongfei Jia, Binqiang Zhao, and Hao Zhang. 2021. 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics. arXiv:2011.09127 [cs.CV] https://arxiv.org/abs/2011.09127

    Pith/arXiv arXiv 2021

  11. [11]

    Bermano, Gal Chechik, and Daniel Cohen-Or

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. 2022. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. arXiv:2208.01618 [cs.CV] https://arxiv.org/abs/2208.01618

    Pith/arXiv arXiv 2022

  12. [12]

    Google. 2025. Nano Banana. https://deepmind.google/models/gemini-image/ flash

    2025

  13. [13]

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. 2024. AnimateDiff: Animate Your Per- sonalized Text-to-Image Diffusion Models without Specific Tuning.International Conference on Learning Representations(2024)

    2024

  14. [14]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. arXiv:2006.11239 [cs.LG] https://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  15. [15]

    Jonathan Ho and Tim Salimans. 2022. Classifier-Free Diffusion Guidance. arXiv:2207.12598 [cs.LG] https://arxiv.org/abs/2207.12598

    Pith/arXiv arXiv 2022

  16. [16]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685

    Pith/arXiv arXiv 2021

  17. [17]

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. 2024. Analyzing and Improving the Training Dynamics of Diffusion Models. arXiv:2312.02696 [cs.CV] https://arxiv.org/abs/2312.02696

    Pith/arXiv arXiv 2024

  18. [18]

    Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. 2022. Revisiting Image Pyramid Structure for High Resolution Salient Object Detection. arXiv:2209.09475 [cs.CV] https://arxiv.org/abs/2209.09475

    Pith/arXiv arXiv 2022

  19. [19]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. 2025. FLUX.1 Kontext: Flow Matching for In-Context ...

    Pith/arXiv arXiv 2025

  20. [20]

    Guang Li, Jintao Song, and Yunxiao Fei. 2025. HomeDiffusion: Zero-Shot Object Customization with Multi-View Representation Learning for Indoor Scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 4698–4706. doi:10.1609/aaai.v39i5.32496

    work page doi:10.1609/aaai.v39i5.32496 2025

  21. [21]

    Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. 2024. Common Diffusion Noise Schedules and Sample Steps are Flawed. arXiv:2305.08891 [cs.CV] https: //arxiv.org/abs/2305.08891

    Pith/arXiv arXiv 2024

  22. [22]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2023. Flow Matching for Generative Modeling. arXiv:2210.02747 [cs.LG] https://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv 2023

  23. [23]

    Open AI. 2025. GPT-Image. https://platform.openai.com/docs/guides/image- generation

    2025

  24. [24]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Trans- formers. arXiv:2212.09748 [cs.CV] https://arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  25. [25]

    Xiangyu Peng, Zangwei Zheng, Chenhui Shen, Tom Young, Xinying Guo, Binluo Wang, Hang Xu, Hongxin Liu, Mingyan Jiang, Wenjun Li, Yuhui Wang, Anbang Ye, Gang Ren, Qianran Ma, Wanying Liang, Xiang Lian, Xiwen Wu, Yuting Zhong, Zhuangyan Li, Chaoyu Gong, Guojun Lei, Leijun Cheng, Limin Zhang, Minghao Li, Ruijie Zhang, Silan Hu, Shijie Huang, Xiaokang Wang, Yu...

    Pith/arXiv arXiv 2025

  26. [26]

    Pinterest. [n. d.]. Pinterest Performance+. https://business.pinterest.com/ pinterest-performance-plus/

  27. [27]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. arXiv:2112.10752 [cs.CV] https://arxiv.org/abs/2112.10752

    Pith/arXiv arXiv 2022

  28. [28]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. arXiv:2208.12242 [cs.CV] https://arxiv.org/abs/ 2208.12242

    Pith/arXiv arXiv 2023

  29. [29]

    Lee, Jonathan Ho, Tim Salimans, David J

    Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-Image Diffusion Models. arXiv:2111.05826 [cs.CV] https://arxiv.org/abs/2111.05826

    Pith/arXiv arXiv 2022

  30. [30]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. arXiv:2011.13456 [cs.LG] https://arxiv.org/abs/2011. 13456

    Pith/arXiv arXiv 2021

  31. [31]

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu

  32. [32]

    RoFormer: Enhanced Transformer with Rotary Position Embedding.arXiv preprint arXiv:2104.09864(2021)

    Pith/arXiv arXiv 2021

  33. [33]

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. 2024. Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction. arXiv:2404.02905 [cs.CV] https://arxiv.org/abs/2404.02905

    Pith/arXiv arXiv 2024

  34. [34]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

    Pith/arXiv arXiv 2025

  35. [35]

    Qilong Wang, Xiaofan Ming, Zhenyi Lin, Jinwen Li, Dongwei Ren, Wangmeng Zuo, and Qinghua Hu. 2025. RoomEditor++: A Parameter-Sharing Diffusion Architecture for High-Fidelity Furniture Synthesis. arXiv:2512.17573 [cs.CV] https://arxiv.org/abs/2512.17573

    arXiv 2025

  36. [36]

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, and Xiaoou Tang. 2018. ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks. arXiv:1809.00219 [cs.CV] https://arxiv.org/ abs/1809.00219

    Pith/arXiv arXiv 2018

  37. [37]

    Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari Morcos, Ali Farhadi, and Ludwig Schmidt. 2023. Stable and low-precision training for large-scale vision-language models. arXiv:2304.13013 [cs.LG] https://arxiv.org/abs/2304. 13013

    Pith/arXiv arXiv 2023

  38. [38]

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

    Pith/arXiv arXiv 2025

  39. [39]

    Keming Wu, Sicong Jiang, Max Ku, Ping Nie, Minghao Liu, and Wenhu Chen

  40. [40]

    arXiv:2509.26346 [cs.CV] https://arxiv.org/abs/2509.26346

    EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing. arXiv:2509.26346 [cs.CV] https://arxiv.org/abs/2509.26346

    arXiv

  41. [41]

    Katherine Xu, Lingzhi Zhang, and Jianbo Shi. 2025. Good Seed Makes a Good Crop: Discovering Secret Seeds in Text-to-Image Diffusion Models. arXiv:2405.14828 [cs.CV] https://arxiv.org/abs/2405.14828

    Pith/arXiv arXiv 2025

  42. [42]

    Jinze Yang, Haoran Wang, Zining Zhu, Chenglong Liu, Meng Wymond Wu, and Mingming Sun. 2024. VIP: Versatile Image Outpainting Empowered by Multimodal Large Language Model. arXiv:2406.01059 [cs.CV] https://arxiv.org/ abs/2406.01059

    Pith/arXiv arXiv 2024

  43. [43]

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. 2023. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. arXiv:2308.06721 [cs.CV] https://arxiv.org/abs/2308.06721

    Pith/arXiv arXiv 2023

  44. [44]

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vi- jay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu

  45. [45]

    arXiv:2206.10789 [cs.CV] https://arxiv.org/abs/2206.10789 Pinterest Canvas: Large-Scale Image Generation at Pinterest KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea

    Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. arXiv:2206.10789 [cs.CV] https://arxiv.org/abs/2206.10789 Pinterest Canvas: Large-Scale Image Generation at Pinterest KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea

    Pith/arXiv arXiv 2026

  46. [46]

    Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji

    Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. 2024. JeDi: Joint-Image Diffusion Models for Finetuning- Free Personalized Text-to-Image Generation. arXiv:2407.06187 [cs.CV] https: //arxiv.org/abs/2407.06187

    Pith/arXiv arXiv 2024

  47. [47]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. arXiv:2302.05543 [cs.CV] https: //arxiv.org/abs/2302.05543

    Pith/arXiv arXiv 2023

  48. [48]

    Sijie Zhao, Jing Cheng, Yaoyao Wu, Hao Xu, and Shaohui Jiao. 2025. DreamPainter: Image Background Inpainting for E-commerce Scenarios. arXiv:2508.02155 [cs.CV] https://arxiv.org/abs/2508.02155 Appendix A Details of Training Dataset Multi-view product.We curate a multi-view product dataset con- sisting of sets of images from the same item, sourced from the...

    Pith/arXiv arXiv 2025