Pith. sign in

REVIEW 2 major objections 3 minor 39 references

RGBA layers, not pixels, become the atomic unit of image generation and editing.

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

2026-08-05 04:49 UTC pith:EQRSAG7O

load-bearing objection Solid layer-decomposition system with a real architecture contribution; the headline editability claim is asserted but not directly tested. the 2 major comments →

arxiv 2608.03971 v1 pith:EQRSAG7O submitted 2026-08-04 cs.CV

UniWorld-Design: From Pixel Generation to Layer-Native Design

Zongjian Li , Zhiyuan Yan , Chenxu Bai , Chen Chen , Haoxiang Sun , Shaodong Wang , Feize Wu , Shenghai Yuan
show 4 more authors
Bin Lin Zheyuan Liu Yuwei Niu Li Yuan
This is my paper
classification cs.CV
keywords layer-native designRGBA layerstext-to-image generationimage-to-layer decompositioninstruction-controlled layeringcomplete semantic objectsdiffusion transformerprogressive distillation
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.

UniWorld-Design argues that the right unit for image generation and editing is not the pixel but the semantic RGBA layer: pixels say how an image is rendered, layers say how it is created, understood, and edited. To make this concrete, the paper builds two models. T2RGBA turns a text prompt directly into a standalone transparent asset, and I2L turns a finished image, a global instruction, and per-layer prompts into an ordered stack of complete semantic RGBA layers, supporting top-level decomposition, recursive decomposition, and targeted extraction. The central wager is that training on complete designer PSD layers, including content hidden behind occluders, produces layers that stay usable when moved or removed, unlike visible-pixel segmentations. On the Crello benchmark, I2L lowers per-layer RGB L1 error by 37% and raises Alpha Soft IoU by 34% relative to Qwen-Image-Layered; T2RGBA reaches the highest mean CLIP Score (33.03) among the compared text-to-RGBA systems. If the claim holds, it gives external design agents a persistent, language-addressable state to plan and execute edits in.

Core claim

The paper's central claim is that image generation should move 'from flat pixel synthesis to structured visual composition,' with semantic RGBA layers as the atomic units of generation, understanding, and editing. T2RGBA generates a standalone transparent asset from a text prompt; I2L takes a finished image, a global instruction, and per-layer prompts and jointly outputs an ordered stack of complete semantic RGBA layers on a shared canvas, supporting top-level, recursive, and targeted extraction. The decisive choice is training I2L on complete PSD-derived semantic layers, so each layer keeps content hidden behind occluders instead of being a visible-pixel cutout. On 512 Crello designs, I2L r

What carries the argument

The load-bearing mechanism is the Layer–Instruction Binding MMDiT (LIB-MMDiT), a diffusion transformer that jointly generates N aligned RGBA layers. Two modifications carry the argument: a layer–instruction binding attention mask, in which the global task instruction broadcasts to all image queries while each per-layer prompt is visible only to that layer's image tokens, and layer-indexed 3D rotary positions, which reuse the transformer's constant first rotary coordinate as a layer index so identical canvas coordinates stay aligned across layers while layer order is preserved. Supporting machinery includes semantic layer trees built from designer-authored PSD documents (which preserve occlud

Load-bearing premise

The load-bearing premise is that a model trained on complete designer-authored layers, where hidden content is visible in the source file, will keep its layers visually coherent when a user actually moves or deletes one; the paper measures editability only through proxy metrics and never performs such an edit.

What would settle it

Take a set of predicted I2L stacks, shift or delete the frontmost layer, composite the remaining layers, and compare the revealed region with the ground-truth complete layer: if the revealed area shows transparency, misaligned color, or the occluder's residual rather than the occluded object, the complete-object claim is refuted. Running this on the same 512 Crello designs the paper's fidelity metrics come from would settle it directly.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • I2L layers become persistent, language-addressable object states: an external agent can decompose at a chosen granularity, recursively split any returned layer, and edit or recompose layers without re-inferring structure from pixels.
  • Because I2L is trained on complete objects, moving or removing an occluding layer should reveal the occluded content rather than a hole, provided the learned complete-object behavior transfers to test images.
  • T2RGBA and I2L close a loop: generated assets can be inserted into a design, the composite can be decomposed back into layers, and selected layers can be refined and recomposed, enabling multi-turn, tool-based design workflows.
  • If the benchmark numbers generalize, layer decomposition quality on real-world design documents is no longer the bottleneck; the remaining burdens are alpha-boundary cleanliness and dense or Chinese typography, which the paper identifies as limitations.

Where Pith is reading between the lines

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

  • The paper does not directly test the editability claim by moving or removing a predicted layer and compositing the result; its editability metrics are proxies. A direct move-or-remove experiment on the same Crello stacks would settle whether complete-object supervision transfers.
  • The same instruction-addressable architecture could be pointed at natural images rather than graphic designs if training data with ground-truth occluded content, such as multi-view or video frames, were available; the method itself does not depend on PSD files.
  • Alpha-boundary quality is the clear next lever: adding a boundary-focused loss or post-processing during distillation could lift the one dimension where I2L trails the baseline and would also improve downstream compositing.
  • If the layer-native space is adopted as an agent interface, evaluation should shift from pixel-fidelity metrics to task-level outcomes, such as whether an agent can complete an edit by extracting, modifying, and recomposing layers, and whether users judge the result editable.

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

2 major / 3 minor

Summary. This paper proposes UniWorld-Design, a framework that treats semantic RGBA layers as native units for generation and decomposition. Two models are presented: Text-to-RGBA (T2RGBA), which generates standalone transparent assets from text, and Image-to-Layer (I2L), which takes a composite image plus a global instruction and per-layer prompts and jointly emits ordered, complete semantic RGBA layers. The I2L architecture uses layer–instruction binding attention and layer-indexed rotary positions (LIB-MMDiT), and training data are derived from designer-authored PSD documents so that occluded content is preserved in target layers. Post-training uses progressive distillation and DiffusionNFT with task-specific rewards. On 512 Crello designs, I2L reports a 37% relative reduction in per-layer RGB L1 (0.2014 to 0.1264) and a 34% relative improvement in Alpha Soft IoU (0.5454 to 0.7325) over Qwen-Image-Layered, along with improved reference-free editability proxies and a higher VLM total score, but a lower alpha-cleanliness score. T2RGBA reports the highest mean CLIP Score among LayerDiffuse, OmniAlpha, and itself, while OmniAlpha leads on FID and white-composite LPIPS.

Significance. The paper addresses a timely and useful problem: making generative image output editable in a layer-native space. The PSD-derived training pipeline is a genuine strength, because it uses real designer-authored layers and preserves occluded content without generating synthetic layer stacks from the same model, thus avoiding a clear source of circularity. The reported quantitative gains are internally consistent with the tables, and the evaluation is multi-metric with standard deviations. However, the central value proposition — that layers remain usable when moved or removed because they are complete semantic objects — is asserted but not directly tested. The current evidence is indirect: per-layer fidelity to complete references, reference-free proxy metrics, and a VLM background-inpainting dimension that the paper itself says is not isolated. If a direct edit experiment were supplied, the contribution would be solid; without it, the core claim rests on inference from training data and proxies.

major comments (2)
  1. [Abstract; §3.4; §3.5; §4.2] The central claim that layers 'stay usable when moved or removed' is load-bearing, but no experiment in Section 4 moves, removes, or recomposes a predicted layer. Table 1 scores per-layer fidelity against complete reference layers after DTW alignment; Table 2 uses reference-free blank/glaze/feature-distribution proxies; Table 3's 'background inpainting' is the only edit-like proxy, and Section 4.2 explicitly states that the effect of complete-object supervision is not isolated. Please add a direct experiment: translate or delete a predicted layer, recompose, and compare against the corresponding reference edit (composite error, VLM/human judged plausibility), ideally with a visible-pixel-partition baseline, to support the causal claim in the abstract and Section 3.4.
  2. [Eq. (6) and Table 1] The I2L post-training reward is a weighted average of RGBA L1 and AlexNet LPIPS against reference layers. The headline evaluation metrics in Table 1 are per-layer RGB L1 and Alpha Soft IoU. The post-training objective is therefore a near-neighbor of the evaluation metric, so the reported 37% and 34% improvements partly reflect optimization of a quantity closely aligned with the evaluation itself. This does not invalidate the result, but it weakens the independent evidentiary value of those numbers for layer usability. Please report at least one metric not aligned with the reward (e.g., a user study or a move/remove composite test) and explicitly acknowledge the reward–metric overlap.
minor comments (3)
  1. [Section 1; Section 4.1; Table 2] Typographical issues: 'difficult' in Section 1, 'official' in Section 4.1, and '1 .34' in Table 2. Please proofread.
  2. [Section 4.3; Abstract] The text correctly says T2RGBA has the highest mean CLIP Score but loses to OmniAlpha on FID and white-composite LPIPS. The abstract's phrasing 'outperforming LayerDiffuse and OmniAlpha' could be read as an overall claim; consider making explicit that the comparison is specifically on CLIP Score.
  3. [Appendix B] The VLM judging protocol names GPT-5.6-terra but does not give the prompt template, temperature, or number of runs. Since the VLM scores are a reported outcome, please include the full judge prompt and sampling details for reproducibility.

Circularity Check

0 steps flagged

No significant circularity. The I2L and T2RGBA results are measured against external baselines and real designer-authored PSD/Crello data; the only mild concern is a reward/evaluation overlap and the untested move/remove editability claim, which is an evidence gap rather than a circular reduction.

full rationale

The paper's empirical chain is self-contained against external benchmarks. I2L is trained on designer-authored PSD documents (Section 3.5: 'We instead use designer-authored PSD documents, where layer content hidden in the final composite remains available in the source file'), and evaluated on 512 held-out Crello designs with no Crello training data (Section 4.1), compared against Qwen-Image-Layered via LayerD alignment metrics. T2RGBA is compared against LayerDiffuse and OmniAlpha on FID, CLIP Score, Alpha MSE, SAD and LPIPS. No fitted parameter is renamed as a prediction: the DiffusionNFT rewards in Eq. (6) use reconstruction losses and an MLLM score, while the evaluation CLIP Score is computed with CLIP ViT-B/32 (Appendix B), distinct from the Qwen3-VL embedding similarity used in the reward. Self-citations exist (e.g., UniWorld-v2 [9] for MLLM implicit feedback), but they supply a training technique rather than a load-bearing uniqueness or equivalence argument, so they do not make the derivation circular. The paper's own limitation note in Section 4.2 ('we do not isolate that effect here') and the absence of a direct move/remove experiment support the skeptic's concern that the 'layers stay usable when moved or removed' claim is under-evidenced; however, under-evidence is a correctness/validity risk, not a circularity of the kind this pass scores.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claims rest on standard diffusion/flow frameworks, a set of training and evaluation assumptions about layer completeness and metrics, and a few hand-tuned reward weights. No new physical entities are introduced.

free parameters (5)
  • Reward weights lambda_Q, lambda_M (T2RGBA) = not disclosed
    Equation 6: weights for CLIP-style and MLLM reward heads; chosen by hand and affect final T2RGBA quality.
  • Reward weights lambda_1, lambda_P (I2L) = equal weights (stable configuration)
    Equation 6: L1 and LPIPS terms combined with equal weight; this choice affects post-trained I2L outputs.
  • Number of distillation steps = 8
    Section 3.6: the student is distilled to eight steps, which sets the evaluation schedule and affects output quality.
  • Number of target layers N = 4 in evaluation
    Section 4.1: fixed to match Qwen-Image-Layered; metrics are sensitive to N.
  • Crello evaluation seed = fixed but unspecified
    Section 4.1: 512 fixed-seed designs; the seed is not disclosed, so the selection is not independently reproducible.
axioms (6)
  • standard math Rectified flow interpolation (Eq. 1) is a valid generative framework.
    Preliminaries, Section 3.1: builds on Lipman et al. and Liu et al.
  • standard math DiffusionNFT objective (Eq. 2) improves flow generations.
    Preliminaries, Section 3.1: taken from Zheng et al. 2026.
  • domain assumption Extending the VAE to 4 channels and aligning RGB latents with a frozen teacher preserves composite fidelity.
    Section 3.2, Eq. 3: the RGB-latent alignment term is assumed sufficient to prevent distribution shift.
  • domain assumption PSD-derived semantic layer trees, grouped with VLM assistance, provide correct ground-truth 'complete semantic objects'.
    Section 3.5: relies on the grouping and hierarchy extraction from designer-authored PSDs.
  • domain assumption The evaluation metrics (RGB L1 with reference-alpha support, Alpha Soft IoU, editability proxies) measure real editing utility.
    Section 4.2 and Appendix B: these are proxy metrics; no direct moved/removed test.
  • domain assumption The pretrained RGB autoencoder's latent space is a reasonable base for RGBA layers.
    Section 3.2: channel extension with zero-initialized alpha filters assumes minimal disruption.

pith-pipeline@v1.3.0-daily-deepseek · 14852 in / 13856 out tokens · 130942 ms · 2026-08-05T04:49:07.738638+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of UniWorld-Design: From Pixel Generation to Layer-Native Design." pith.science (2026). https://pith.science/paper/EQRSAG7O

@misc{pith2026260803971,
  author       = {Pith},
  title        = {Pith review of: UniWorld-Design: From Pixel Generation to Layer-Native Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQRSAG7O}},
  note         = {Machine review of arXiv:2608.03971}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce UniWorld-Design, a framework that redefines image generation from flat pixel synthesis to structured visual composition, with semantic RGBA layers as the atomic units of generation, understanding, and editing. Our key insight is that pixels define how an image is rendered, whereas layers define how an image is created, understood, and edited. Just as human designers create and manipulate visual content through layers rather than raw pixels, UniWorld-Design equips multimodal generative models with a layer-native design space. UniWorld-Design comprises two models. The Text-to-RGBA (T2RGBA) model generates standalone RGBA assets directly from text. The Image-to-Layer (I2L) model conditions on a finished image, a global instruction and per-layer prompts, and jointly produces ordered, complete semantic RGBA layers. Its instruction interface supports top-level decomposition, recursive decomposition and targeted extraction, making layering an instruction-addressable operation for agentic editing. Because I2L learns complete semantic objects rather than visible-pixel partitions, its layers stay usable when moved or removed. On the Crello benchmark, I2L reduces per-layer RGB L1 error by 37% and achieves a 34% relative improvement in Alpha Soft IoU over Qwen-Image-Layered. Separately, T2RGBA achieves the highest CLIP Score, outperforming LayerDiffuse and OmniAlpha.

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

39 extracted references · 26 canonical work pages · 2 internal anchors

  1. [1]

    Qwen3-vl technical report

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

    Pith/arXiv arXiv 2025

  2. [2]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 18392– 18402, 2023

    work page 2023

  3. [3]

    Prismlayers: Open data for high-quality multi-layer transparent image generative models

    Junwen Chen, Heyang Jiang, Yanbin Wang, Keming Wu, Ji Li, Chao Zhang, Keiji Yanai, Dong Chen, and Yuhui Yuan. Prismlayers: Open data for high-quality multi-layer transparent image generative models. arXiv preprint arXiv:2505.22523, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Diffedit: Diffusion-based semantic image editing with mask guidance

    Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In International Conference on Learning Representations (ICLR) , 2023

    work page 2023

  5. [5]

    Sensenova-u1: Unifying multimodal understanding and generation with NEO-unify architecture

    Haiwen Diao, Penghao Wu, Hanming Deng, Jiahao Wang, Shihao Bai, Silei Wu, Weichen Fan, Wenjie Ye, Wenwen Tong, Xiangyu Fan, et al. Sensenova-u1: Unifying multimodal understanding and generation with NEO-unify architecture. arXiv preprint arXiv:2605.12500 , 2026

    Pith/arXiv arXiv 2026

  6. [6]

    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, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML) , 2024

    work page 2024

  7. [7]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS) , 2017

    work page 2017

  8. [8]

    A closed-form solution to natural image matting

    Anat Levin, Dani Lischinski, and Yair Weiss. A closed-form solution to natural image matting. IEEE Transactions on Pattern Analysis and Machine Intelligence , 30(2):228–242, 2008

    work page 2008

  9. [9]

    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

  10. [10]

    Uniworld: 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: High-resolution semantic encoders for unified visual understanding and generation. arXiv preprint arXiv:2506.03147 , 2025

    Pith/arXiv arXiv 2025

  11. [11]

    SDXL-Lightning: Progressive adversarial diffusion distillation

    Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929 , 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In International Conference on Learning Representations (ICLR) , 2023

    work page 2023

  13. [13]

    Omnipsd: Layered psd generation with diffusion transformer

    Cheng Liu, Yiren Song, Haofan Wang, and Mike Zheng Shou. Omnipsd: Layered psd generation with diffusion transformer. arXiv preprint arXiv:2512.09247 , 2025

    arXiv 2025

  14. [14]

    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. In International Conference on Learning Representations (ICLR) , 2023

    work page 2023

  15. [15]

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

    Yuwei Niu, Munan Ning, Mengren Zheng, Weiyang Jin, Bin Lin, Peng Jin, Jiaqi Liao, Chaoran Feng, Kunpeng Ning, Bin Zhu, et al. WISE: A world knowledge-informed semantic evaluation for text-to-image generation. arXiv preprint arXiv:2503.07265 , 2025

    Pith/arXiv arXiv 2025

  16. [16]

    pix2gestalt: Amodal segmentation by synthesizing wholes

    Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  17. [17]

    Art: Anonymous region transformer for variable multi-layer transparent image generation

    Yifan Pu, Yiming Zhao, Zhicong Tang, Ruihong Yin, Haoxing Ye, Yuhui Yuan, Dong Chen, Jianmin Bao, Sirui Zhang, Yanbin Wang, Lin Liang, Lijuan Wang, Ji Li, Xiu Li, Zhouhui Lian, et al. Art: Anonymous region transformer for variable multi-layer transparent image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2025. 16

    work page 2025

  18. [18]

    Highly accurate dichotomous image segmentation

    Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision (ECCV) , 2022

    work page 2022

  19. [19]

    Qwen3.5-9B

    Qwen Team. Qwen3.5-9B. Hugging Face model card, 2026. URL https://huggingface.co/Qwen/Qwen3.5-9B

    work page 2026

  20. [20]

    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, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML) , 2021

    work page 2021

  21. [21]

    Stable-Layers: Fine-Tuning Image Layer Decomposition Models with VLM-Scored Reinforcement Learning

    Ciara Rowles, Reshinth Adithyan, Nikhil Pinnaparaju, Vikram Voleti, and Mark Boss. Stable-layers: Fine-tuning image layer decomposition models with vlm-scored reinforcement learning. arXiv preprint arXiv:2605.30257 , 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [22]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR) , 2022

    work page 2022

  23. [23]

    Layerd: Decomposing raster graphic designs into layers

    Tomoyuki Suzuki, Kang-Jun Liu, Naoto Inoue, and Kota Yamaguchi. Layerd: Decomposing raster graphic designs into layers. In IEEE/CVF International Conference on Computer Vision (ICCV) , 2025

    work page 2025

  24. [24]

    MULAN: A Multi Layer Annotated Dataset for Controllable Text-to-Image Generation

    Petru-Daniel Tudosiu, Yongxin Yang, Shifeng Zhang, Fei Chen, Steven McDonagh, Gerasimos Lampouras, Ig- nacio Iacobacci, and Sarah Parisot. Mulan: A multi layer annotated dataset for controllable text-to-image generation. arXiv preprint arXiv:2404.02790 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Genartist: Multimodal llm as an agent for unified image generation and editing

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. In Advances in Neural Information Processing Systems (NeurIPS) , volume 37, 2024

    work page 2024

  26. [26]

    Alphavae: Unified end-to-end rgba image reconstruction and generation with alpha-aware representation learning

    Zile Wang, Hao Yu, Jiabo Zhan, and Chun Yuan. Alphavae: Unified end-to-end rgba image reconstruction and generation with alpha-aware representation learning. arXiv preprint arXiv:2507.09308 , 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Canvasvae: Learning to generate vector graphic documents

    Kota Yamaguchi. Canvasvae: Learning to generate vector graphic documents. In IEEE/CVF International Conference on Computer Vision (ICCV) , 2021

    work page 2021

  28. [28]

    Unified multimodal models as auto-encoders

    Zhiyuan Yan, Kaiqing Lin, Zongjian Li, Junyan Ye, Hui Han, Haochen Wang, Zhendong Wang, Bin Lin, Hao Li, Xinyan Xiao, Jingdong Wang, Haifeng Wang, and Li Yuan. Unified multimodal models as auto-encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 41903–41912, June 2026

    work page 2026

  29. [29]

    Agent banana: High-fidelity image editing with agentic thinking and tooling

    Ruijie Ye, Jiayi Zhang, Zhuoxin Liu, Zihao Zhu, Siyuan Yang, Li Li, et al. Agent banana: High-fidelity image editing with agentic thinking and tooling. arXiv preprint arXiv:2602.09084 , 2026

    arXiv 2026

  30. [30]

    Ni, Jingren Zhou, Junyang Lin, and Chenfei Wu

    Shengming Yin, Zekai Zhang, Zecheng Tang, Kaiyuan Gao, Xiao Xu, Kun Yan, Jiahao Li, Yilei Chen, Yuxiang Chen, Heung-Yeung Shum, Lionel M. Ni, Jingren Zhou, Junyang Lin, and Chenfei Wu. Qwen-image-layered: Towards inherent editability via layer decomposition. arXiv preprint arXiv:2512.15603 , 2025

    arXiv 2025

  31. [31]

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In Advances in Neural Information Processing Systems (NeurIPS) , 2024

    work page 2024

  32. [32]

    Freeman, and Taesung Park

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2024

    work page 2024

  33. [33]

    Omnialpha: Aligning transparency-aware generation via multi-task unified reinforcement learning

    Hao Yu, Jinglin Wang, Jiabo Zhan, Rui Chen, Zile Wang, Huaisong Zhang, Hongyu Li, Xinrui Chen, Yongxian Wei, and Chun Yuan. Omnialpha: Aligning transparency-aware generation via multi-task unified reinforcement learning. arXiv preprint arXiv:2511.20211 , 2025

    Pith/arXiv arXiv 2025

  34. [34]

    ReDesign: Recovering editable design structures from images via agentic decomposition

    Jooyeol Yun, Jintae Park, Hyesu Lim, Junha Hyung, Hyungjin Chung, and Jaegul Choo. ReDesign: Recovering editable design structures from images via agentic decomposition. In European Conference on Computer Vision (ECCV), 2026

    work page 2026

  35. [35]

    Transparent image layer diffusion using latent transparency

    Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 17

    work page 2018

  37. [37]

    Diffusionnft: Online diffusion reinforcement with forward process

    Kaiwen Zheng, Huayu Chen, Haotian Ye, Haoxiang Wang, Qinsheng Zhang, Kai Jiang, Hang Su, Stefano Ermon, Jun Zhu, and Ming-Yu Liu. Diffusionnft: Online diffusion reinforcement with forward process. In International Conference on Learning Representations (ICLR) , 2026. Oral

    work page 2026

  38. [38]

    Bilateral reference for high-resolution dichotomous image segmentation

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research , 3:9150038, 2024

    work page 2024

  39. [39]

    Semantic amodal segmentation

    Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2017. 18 A Contributions Contributors: Zongjian Li, Chenxu Bai, Chen Chen, Haoxiang Sun, Shaodong Wang, Feize Wu, Shenghai Yuan, Bin Lin, Zheyuan Liu, Yuwei Niu. ProjectLeads : Zhiyuan Yan, Li...

    work page 2017