pith. sign in

arxiv: 2605.30257 · v1 · pith:M3EGOZNBnew · submitted 2026-05-28 · 💻 cs.CV

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

Ciara Rowles , Reshinth Adithyan , Nikhil Pinnaparaju , Vikram Voleti , Mark Boss This is my paper

Pith reviewed 2026-06-29 08:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords layer decompositionreinforcement learningvision-language modelsfine-tuningimage editingGRPOCrello dataset
0
0 comments X

The pith

Stable-Layers fine-tunes layer decomposition models with VLM-scored reinforcement learning and no paired supervision.

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

The paper presents a reinforcement learning approach that starts from a pretrained layer decomposition model and improves it solely through feedback from a vision-language model. It solves the problem of narrow score ranges by using a two-stage pipeline: first scoring each candidate decomposition on five edit-centric criteria, then calibrating all candidates together in a grid view to restore variance. With this reward signal the method runs Flow-GRPO plus LoRA adaptation, sampling multiple decompositions per image and updating the policy from group-relative advantages. The result is decompositions that separate layers more cleanly, contain fewer blank or artifact-laden outputs, and show lower per-layer reconstruction error on the Crello dataset.

Core claim

Starting from Qwen-Image-Layered, Stable-Layers applies Flow-GRPO with LoRA adaptation, sampling multiple candidate decompositions per image, scoring them with a VLM through a two-stage pipeline of per-sample criteria scoring followed by grid-based side-by-side calibration, and optimising the policy from group-relative advantages, yielding stronger layer separation, fewer blank or artifact-heavy layers, and lower per-layer reconstruction error on Crello compared with the base model.

What carries the argument

The two-stage VLM evaluation pipeline that pairs structured per-sample scoring across five edit-centric criteria with a grid-based calibration step in which the VLM re-scores all candidates side-by-side.

If this is right

  • Decompositions exhibit stronger layer separation than the base model.
  • The number of blank or artifact-heavy layers decreases.
  • Per-layer reconstruction error drops on the Crello dataset.
  • Fine-tuning becomes possible without any paired ground-truth decompositions.

Where Pith is reading between the lines

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

  • The same calibrated VLM reward construction could be tested on other generative tasks where human preference data is expensive but visual quality is easy to judge.
  • If the calibration step proves essential, future VLM-as-reward pipelines may need an explicit group comparison stage rather than isolated scoring.
  • The method could be re-run with different base VLMs to test whether the quality of the reward model itself limits further gains.

Load-bearing premise

The two-stage VLM evaluation pipeline supplies a sufficiently reliable and high-variance reward signal that enables effective policy improvement via Flow-GRPO.

What would settle it

Running the identical training loop but omitting the grid-based calibration step and checking whether within-group score variance collapses to the point that policy updates become negligible, or measuring whether the reported gains on layer separation and reconstruction error vanish on a fresh held-out image set.

Figures

Figures reproduced from arXiv: 2605.30257 by Ciara Rowles, Mark Boss, Nikhil Pinnaparaju, Reshinth Adithyan, Vikram Voleti.

Figure 1
Figure 1. Figure 1: Stable-Layers. We finetune a layer decomposition model using Flow-GRPO and a VLM judge, and improve layerization without relying on paired data. The resulting layers have improved consistency, separation and handle in-painting of occluded areas better. Abstract We present Stable-Layers, a reinforcement learning framework that eliminates the need for paired supervision by fine-tuning a pretrained layer deco… view at source ↗
Figure 2
Figure 2. Figure 2: Stable-Layers training pipeline. Sample G candidates, score with the two-phase VLM reward, replay with GRPO updates to LoRA parameters. variance, preventing the clipped surrogate from constraining overconfident positive-advantage updates. GRPO-Guard addresses this with two corrections: (i) RatioNorm, which standardizes log ρg per denoising step so that the ratio distribution is centered near 1 with uniform… view at source ↗
Figure 4
Figure 4. Figure 4: Held-out evaluation metrics. Three auto￾mated metrics on 480 LAION-Aesthetics [22] images across training. Top: bad layers per decomposition (blank + glaze; lower is better) fall from ∼1.65 to ∼0.4. Middle: feature distribution evenness (higher is better) rises from ∼0.53 to ∼0.73. Bottom: layer 0 inpainting quality (higher is better) rises from ∼0.38 to ∼0.62. Bands show ±1σ across the set. 6.2 Qualitativ… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on held-out images. Base model (Qwen Image Layered, top of each pair) vs. Stable-Layers-fine-tuned model (Stable-Layers, bottom). Columns show the input, the composite, and individual layers on white backgrounds. The fine-tuned model produces plausible background inpainting on layer 0 and isolates distinct semantic elements across foreground layers, where the base model leaves layer … view at source ↗
Figure 6
Figure 6. Figure 6: Extended qualitative gallery. Layer decompositions from the Stable-Layers-fine-tuned model on a diverse set of held-out inputs. Each row shows the reconstructed composite and the individual layers composited onto white backgrounds. Examples illustrate consistent behaviour across subject matter: clean isolation of foreground objects (corgi, swan, deer, jacket), plausible background inpainting where foregrou… view at source ↗
Figure 7
Figure 7. Figure 7: Calibration ablation (additional Layer 0 metrics). Layer 0 combined quality (left) and edge-density sharpness (right) over training, comparing the full two-phase reward with grid calibration against Phase 1 individual scoring alone. Both metrics show the calibrated run maintaining a small but consistent lead from approximately step 120 onward, corroborating the SSIM result reported in [PITH_FULL_IMAGE:fig… view at source ↗
Figure 8
Figure 8. Figure 8: Mean VLM reward during training. Phase 2 calibrated reward per step (light) and rolling average (dark). The mean reward rises over the first ∼100 steps as the policy eliminates the worst failure modes, then plateaus around 0.83–0.85 with high per-step variance. Under GRPO’s within-group normalisation (Equation (3)), learning requires only relative discrimination among group members, not rising absolute sco… view at source ↗
read the original abstract

We present Stable-Layers, a reinforcement learning framework that eliminates the need for paired supervision by fine-tuning a pretrained layer decomposition model using only feedback from a vision-language model (VLM). Starting from Qwen-Image-Layered, we apply Flow-GRPO with LoRA adaptation, sampling multiple candidate decompositions per image, scoring them with a VLM, and optimising the policy from group-relative advantages. The key challenge lies in designing a reliable reward signal: VLMs scoring samples in isolation tend to compress their judgements into a narrow band, leaving GRPO with little within-group variance to learn from. We address this with a two-stage evaluation pipeline that pairs structured per-sample scoring across five edit-centric criteria with a grid-based calibration step in which the VLM re-scores all candidates side-by-side. Stable-Layers produces decompositions with stronger layer separation, fewer blank or artifact-heavy layers, and lower per-layer reconstruction error on the Crello dataset compared to the base model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes Stable-Layers, a reinforcement learning framework for fine-tuning a pretrained image layer decomposition model (Qwen-Image-Layered) using only VLM feedback via Flow-GRPO with LoRA. It introduces a two-stage VLM scoring pipeline (per-sample scoring on five criteria followed by grid-based side-by-side calibration) to address score compression and low variance in rewards. The method is claimed to produce better layer decompositions with stronger separation, fewer artifacts, and lower reconstruction error on the Crello dataset compared to the base model.

Significance. If the VLM-based reward signal proves reliable, this approach could enable effective fine-tuning of decomposition models without requiring paired supervision data, which is often scarce. However, the current manuscript provides no quantitative results or validation of the reward, limiting the ability to assess its significance.

major comments (2)
  1. [Abstract] Abstract: the abstract states that Stable-Layers produces decompositions with stronger layer separation, fewer blank or artifact-heavy layers, and lower per-layer reconstruction error on the Crello dataset, but supplies no quantitative results, error bars, ablation details, or dataset statistics to support these claims.
  2. [Abstract] Abstract: the central claim relies on the two-stage VLM evaluation pipeline supplying a reliable reward signal, but no correlation analysis is provided between the VLM scores and human expert ratings or objective proxies such as layer-wise IoU against Crello ground-truth masks.
minor comments (1)
  1. The description of the five edit-centric criteria could be expanded for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for stronger quantitative support in the abstract and validation of the VLM-based reward. We will revise the manuscript accordingly to address these points while preserving the core contributions of the two-stage scoring pipeline and Flow-GRPO fine-tuning approach.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the abstract states that Stable-Layers produces decompositions with stronger layer separation, fewer blank or artifact-heavy layers, and lower per-layer reconstruction error on the Crello dataset, but supplies no quantitative results, error bars, ablation details, or dataset statistics to support these claims.

    Authors: We agree the abstract should be more self-contained with quantitative backing. In the revision we will insert concrete metrics drawn from the experimental section (e.g., mean per-layer reconstruction error reduction, layer-separation scores, and Crello dataset statistics) together with error bars from repeated runs and a brief reference to the ablation studies already present in the main text. revision: yes

  2. Referee: [Abstract] Abstract: the central claim relies on the two-stage VLM evaluation pipeline supplying a reliable reward signal, but no correlation analysis is provided between the VLM scores and human expert ratings or objective proxies such as layer-wise IoU against Crello ground-truth masks.

    Authors: We accept that explicit validation of the reward signal strengthens the central claim. The revised manuscript will include a new subsection reporting layer-wise IoU correlations against Crello ground-truth masks and, where feasible, a small-scale human rating study. If resource constraints limit the human study, we will clearly state this as a limitation while still providing the IoU analysis. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper's core method applies Flow-GRPO to a pretrained layer decomposition model using reward signals generated by an external VLM through a two-stage per-sample and grid-calibration pipeline. This reward is independent of any quantities defined or fitted inside the decomposition model itself, and the reported improvements (layer separation, reconstruction error on Crello) are measured against the base model on held-out data rather than being tautological to the training objective. No self-definitional loops, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation; the approach is externally grounded in VLM feedback and dataset evaluation.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract; full implementation details unavailable. The primary unverified premise is that VLM judgments on the chosen criteria correlate with human-preferred layer quality.

free parameters (2)
  • five edit-centric criteria
    Chosen by authors to structure VLM scoring; exact definitions and weighting not specified in abstract.
  • grid calibration parameters
    Number of candidates per group and grid layout are design choices that affect score variance.
axioms (1)
  • domain assumption Vision-language models can produce consistent and informative scores for image layer decompositions when given structured per-sample criteria plus side-by-side comparison.
    This assumption underpins the entire reward signal and is stated as the key challenge the paper addresses.

pith-pipeline@v0.9.1-grok · 5721 in / 1314 out tokens · 53614 ms · 2026-06-29T08:23:57.520892+00:00 · methodology

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

45 extracted references · 8 canonical work pages · 2 internal anchors

  1. [1]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  2. [2]

    Referring layer decomposition.arXiv preprint arXiv:2602.19358, 2026

    Fangyi Chen, Yaojie Shen, Lu Xu, Ye Yuan, Shu Zhang, Yulei Niu, and Longyin Wen. Referring layer decomposition.arXiv preprint arXiv:2602.19358, 2026

    work page arXiv 2026

  3. [3]

    From inpainting to layer decomposition: Repurposing generative inpainting models for image layer decomposition.arXiv preprint arXiv:2511.20996, 2025

    Jingxi Chen, Yixiao Zhang, Xiaoye Qian, Zongxia Li, Cornelia Fermuller, Caren Chen, and Yiannis Aloimonos. From inpainting to layer decomposition: Repurposing generative inpainting models for image layer decomposition.arXiv preprint arXiv:2511.20996, 2025

    work page arXiv 2025

  4. [4]

    Topreward: Token probabilities as hidden zero-shot rewards for robotics.arXiv preprint arXiv:2602.19313, 2026

    Shirui Chen, Cole Harrison, Ying-Chun Lee, Angela Jin Yang, Zhongzheng Ren, Lillian J Ratliff, Jiafei Duan, Dieter Fox, and Ranjay Krishna. Topreward: Token probabilities as hidden zero-shot rewards for robotics.arXiv preprint arXiv:2602.19313, 2026

    work page arXiv 2026

  5. [5]

    MJ-bench: Is your multimodal reward model really a good judge for text-to-image generation?arXiv preprint, 2024

    Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhe Tong, Qing Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. MJ-bench: Is your multimodal reward model really a good judge for text-to-image generation?arXiv preprint, 2024

    2024

  6. [6]

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  7. [7]

    LayerFusion: Harmonized multi-layer text-to-image generation with generative priors.arXiv preprint, 2024

    Yusuf Dalva, Yijun Li, Qing Liu, Nanxuan Zhao, Jianming Zhang, Zhe Lin, and Pinar Yanardag. LayerFusion: Harmonized multi-layer text-to-image generation with generative priors.arXiv preprint, 2024

    2024

  8. [8]

    DPOK: Reinforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. DPOK: Reinforcement learning for fine-tuning text-to-image diffusion models. InNeural Information Processing Systems (NeurIPS), 2023

    2023

  9. [9]

    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. LoRA: Low-rank adaptation of large language models.arXiv preprint, 2021

    2021

  10. [10]

    PSDiffusion: Harmonized multi-layer image generation via layout and appearance alignment

    Dingbang Huang, Wenbo Li, Yifei Zhao, Xinyu Pan, Yanhong Zeng, and Bo Dai. PSDiffusion: Harmonized multi-layer image generation via layout and appearance alignment. InWinter Conference on Applications of Computer Vision (WACV), 2026

    2026

  11. [11]

    Dreamlayer: Simultaneous multi-layer generation via diffusion model

    Junjia Huang, Pengxiang Yan, Jinhang Cai, Jiyang Liu, Zhao Wang, Yitong Wang, Xinglong Wu, and Guanbin Li. Dreamlayer: Simultaneous multi-layer generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3357–3366, 2025

    2025

  12. [12]

    LayerDiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model

    Runhui Huang et al. LayerDiff: Exploring text-guided multi-layered composable image synthesis via layer-collaborative diffusion model. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  13. [13]

    Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge.arXiv preprint arXiv:2501.01197, 2025

    Kyoungkook Kang, Gyujin Sim, Geonung Kim, Donguk Kim, Seungho Nam, and Sunghyun Cho. Layeringdiff: Layered image synthesis via generation, then disassembly with generative knowledge.arXiv preprint arXiv:2501.01197, 2025

    work page arXiv 2025

  14. [14]

    Self-Rewarding Vision-Language Model via Reasoning Decomposition

    Zongxia Li, Wenhao Yu, Chengsong Huang, Rui Liu, Zhenwen Liang, Fuxiao Liu, Jingxi Che, Dian Yu, Jordan Boyd-Graber, Haitao Mi, et al. Self-rewarding vision-language model via reasoning decomposition.arXiv preprint arXiv:2508.19652, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    See-through: Single-image layer decomposition for anime characters.arXiv preprint arXiv:2602.03749, 2026

    Jian Lin, Chengze Li, Haoyun Qin, Kwun Wang Chan, Yanghua Jin, Hanyuan Liu, Stephen Chun Wang Choy, and Xueting Liu. See-through: Single-image layer decomposition for anime characters.arXiv preprint arXiv:2602.03749, 2026

    work page arXiv 2026

  16. [16]

    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.arXiv preprint arXiv:2505.05470, 2025. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint, 2022

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint, 2022

    2022

  18. [18]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint, 2022

    2022

  19. [19]

    Controllable layer decomposition for reversible multi-layer image generation.arXiv preprint, 2025

    Zihao Liu, Zunnan Xu, Shi Shu, Jun Zhou, Ruicheng Zhang, Zhenchao Tang, and Xiu Li. Controllable layer decomposition for reversible multi-layer image generation.arXiv preprint, 2025

    2025

  20. [20]

    Fine-T2I: An open, large-scale, and diverse dataset for high-quality T2I fine-tuning.arXiv preprint, 2026

    Xu Ma, Yitian Zhang, Qihua Dong, and Yun Fu. Fine-T2I: An open, large-scale, and diverse dataset for high-quality T2I fine-tuning.arXiv preprint, 2026

    2026

  21. [21]

    Aligning text-to- image diffusion models with reward backpropagation.arXiv preprint, 2023

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to- image diffusion models with reward backpropagation.arXiv preprint, 2023

    2023

  22. [22]

    LAION- 5B: An open large-scale dataset for training next generation image-text models.Neural Infor- mation Processing Systems (NeurIPS), 35, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION- 5B: An open large-scale dataset for training next generation image-text models.Neural Infor- mation Processing Systems (NeurIPS), 35, 2022

    2022

  23. [23]

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

    2024

  24. [24]

    LayerD: Decomposing raster graphic designs into layers

    Tomoyuki Suzuki, Kang-Jun Liu, Naoto Inoue, and Kota Yamaguchi. LayerD: Decomposing raster graphic designs into layers. InInternational Conference on Computer Vision (ICCV), 2025

    2025

  25. [25]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purber, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. InConference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  26. [26]

    Grpo-guard: Mitigating implicit over-optimization in flow matching via regulated clipping.arXiv preprint, 2025

    Jing Wang, Jiajun Liang, Jie Liu, Henglin Liu, Gongye Liu, Jun Zheng, Wanyuan Pang, Ao Ma, Zhenyu Xie, Xintao Wang, et al. Grpo-guard: Mitigating implicit over-optimization in flow matching via regulated clipping.arXiv preprint, 2025

    2025

  27. [27]

    Self-evolving vision-language models for image quality assessment via voting and ranking.arXiv preprint, 2025

    Wen Wen, Tianwu Zhi, Kanglong Fan, Yang Li, Xinge Peng, Yabin Zhang, Yiting Liao, Junlin Li, and Li Zhang. Self-evolving vision-language models for image quality assessment via voting and ranking.arXiv preprint, 2025

    2025

  28. [28]

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

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint, 2023

    2023

  29. [29]

    Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation

    Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. InAAAI Conference on Artificial Intelligence (AAAI), pages 11269–11277, 2026

    2026

  30. [30]

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

    Jiazheng Xu et al. ImageReward: Learning and evaluating human preferences for text-to-image generation. InNeural Information Processing Systems (NeurIPS), 2023

    2023

  31. [31]

    DanceGRPO: Unleashing GRPO on visual generation.arXiv preprint, 2025

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, and Ping Luo. DanceGRPO: Unleashing GRPO on visual generation.arXiv preprint, 2025

    2025

  32. [32]

    CanvasV AE: Learning to generate vector graphic documents.International Conference on Computer Vision (ICCV), 2021

    Kota Yamaguchi. CanvasV AE: Learning to generate vector graphic documents.International Conference on Computer Vision (ICCV), 2021

    2021

  33. [33]

    Generative image layer decomposition with visual effects

    Jinrui Yang, Qing Liu, Yijun Li, Soo Ye Kim, Daniil Pakhomov, Mengwei Ren, Jianming Zhang, Zhe Lin, Cihang Xie, and Yuyin Zhou. Generative image layer decomposition with visual effects. InConference on Computer Vision and Pattern Recognition (CVPR), pages 7643–7653, 2025

    2025

  34. [34]

    Controllable layered image generation for real-world editing.arXiv preprint, 2026

    Jinrui Yang, Qing Liu, Yijun Li, Mengwei Ren, Letian Zhang, Zhe Lin, Cihang Xie, and Yuyin Zhou. Controllable layered image generation for real-world editing.arXiv preprint, 2026

    2026

  35. [35]

    Qwen-image-layered: Towards inherent editability via layer decomposition.arXiv preprint arXiv:2512.15603, 2025

    Shengming Yin, Zekai Zhang, Zecheng Tang, Kaiyuan Gao, Xiao Xu, Kun Yan, Jiahao Li, Yilei Chen, Yuxiang Chen, Heung-Yeung Shum, et al. Qwen-image-layered: Towards inherent editability via layer decomposition.arXiv preprint arXiv:2512.15603, 2025. 11

    work page arXiv 2025

  36. [36]

    Transparent image layer diffusion using latent trans- parency.ACM Transactions on Graphics (TOG), 2024

    Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent trans- parency.ACM Transactions on Graphics (TOG), 2024

    2024

  37. [37]

    Direct preference optimization of video large multimodal models from language model reward

    Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander G Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguisti...

    2025

  38. [38]

    Workflow-aware structured layer decomposition for illustration production.arXiv preprint, 2026

    Tianyu Zhang, Dongchi Li, Keiichi Sawada, and Haoran Xie. Workflow-aware structured layer decomposition for illustration production.arXiv preprint, 2026. 12 A Algorithm Pseudocode Algorithm 1Stable-Layers Training Step (adapted from Flow-GRPO with GRPO-Guard stabilisation) Require: Policy πθ (LoRA), reference πref, reward model R, group size G, latent dim...

    2026

  39. [39]

    a person, a car, a tree)

    semantic_separation (0-5): Each foreground layer should contain ONE distinct, complete object or semantic element (e.g. a person, a car, a tree). Score 0 if a single object is arbitrarily split across multiple layers or if layers contain random crops/slices of the scene rather than meaningful elements. Score 5 if every foreground layer isolates a complete...

  40. [40]

    Score 0 if layers show a semi-transparent haze, ghosting, colour bleed, or a milky/glazed wash over areas that should be fully transparent

    alpha_cleanliness (0-5): Foreground layers should have crisp, binary-like alpha with clean edges. Score 0 if layers show a semi-transparent haze, ghosting, colour bleed, or a milky/glazed wash over areas that should be fully transparent. Transparent regions must be FULLY transparent with zero colour residue. Score 5 if alpha masks are sharp, edges are cle...

  41. [41]

    Score 0 if the background is blurry, has obvious holes, smeared patches, or copy-paste artifacts where foreground objects were removed

    background_inpainting (0-5): Layer 0 (the background) should look like a plausible complete scene with foreground objects removed and their regions filled in convincingly. Score 0 if the background is blurry, has obvious holes, smeared patches, or copy-paste artifacts where foreground objects were removed. Score 5 if the inpainted regions blend seamlessly...

  42. [42]

    Score 0 if most content is crammed into one layer while others are blank or near-empty

    feature_distribution (0-5): Visual content should be meaningfully spread across layers. Score 0 if most content is crammed into one layer while others are blank or near-empty. Score 5 if layers have a balanced, meaningful distribution of the scene’s content

  43. [43]

    semantic_separation

    content_validity (0-5): Penalize blank, empty, or noise-only layers. Score 0 if most layers are blank or contain only noise/blur. Score 5 if all layers have clear, recognizable content. - total (0-25): Sum of all five scores. Return ONLY valid JSON: {"semantic_separation":X, "alpha_cleanliness":Y, "background_inpainting":Z, "feature_distribution":W, "cont...

  44. [44]

    Limitations

    found that this asymmetry did not degrade final sample quality while substantially reducing training cost. Trajectory replay.For each stored SDE step i, the current policy’s transition meanµθ i is recomputed via a forward pass through the LoRA-adapted transformer with gradients enabled. The KL reference is computed by running the same forward pass with Lo...

  45. [45]

    • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

    2025