Image-to-layer decomposition converts a flattened image into editable RGBA layers, enabling element-level editing in design workflows. Existing diffusion-based systems typically adapt large pretrained text-to-image (T2I) models and introduce RGBA autoencoders or variable-layer architectural modules. We revisit this design choice and ask whether layer decomposition truly requires these heavyweight components. We introduce PixelART, a pixel-space rectified-flow Transformer trained from scratch for image-to-layer (I2L) decomposition. PixelART directly denoises regional RGBA pixel patches using a single-stream multi-modal diffusion Transformer, avoiding RGBA-VAEs, pretrained T2I backbones, and layer-specific decoders. We identify a key property of the task: high-noise timesteps determine layer assignment and coarse layer organization, while low-noise timesteps mainly refine color, alpha, texture, and boundaries. Based on this observation, we propose a terminal-boosted timestep sampling strategy to increase training coverage in the high-noise layer assignment regime. Trained on 4M multi-layer design templates, PixelART achieves state-of-the-art layer decomposition and composite reconstruction on Design-Multi-Layer-Bench and LICA with over 80% fewer parameters, 98% lower latency, and 85% lower memory than the recent Qwen-Image-Layered model. Ablation experiments show that pixel-space x-prediction, terminal-boosted timestep sampling, and data/model scaling are critical, while T2I pretraining brings marginal benefits to the I2L task.
Figures & tables
Figure 1: Comparison with previous Image-to-Layer decomposition systems. Left: ART [ 47 ] and its variant; Middle: Qwen-Image-Layered [ 60 ] ; Right: Ours. Our PixelART is much simpler and trained from scratch.
Figure 2: Image-to-layer decomposition results with our PixelART-H/32.
Figure 3: High-noise timesteps are critical for layer assignment. We compare layer states along the sampling trajectory for checkpoints with and without terminal boosting. The boosted model (1st row) assigns regions correctly at the earliest high-noise stage, whereas the unboosted model (2nd row) makes a persistent layer-assignment error. Later steps mainly refine RGBA details. Green/red boxes mark correct/incorrect assignments.
Figure 4: Terminal-boosted timestep distributions. We mix the base schedule with either a point mass at t=0 or a uniform distribution U(0,τ) to increase the mass in the high-noise region.
Figure 5: User study. Comparison with previous SoTA methods and the commercial system Magic Layers [ 4 ] .
Design-Multi-Layer-Bench
LICA
Pixel PSNR ↑
VLM (GPT/Gem.) ↑
Pixel PSNR ↑
VLM (GPT/Gem.) ↑
Inference Efficiency
Method
T2I Pre.
VAE
#Params (Layout)
Comp.
Alpha
Visual
Comp.
Alpha
Visual
Time (s) ↓
Mem (GB) ↓
CLD [ 36 ]
✓
✓
14B (12+2)
29.79
4.24/4.83
4.50/4.68
28.84
4.21/4.78
4.38/4.62
58.7 (+0.7)
55.6 (+4.2)
Qwen-Image-Layered [ 60 ]
✓
✓
20B
32.46
4.30/4.73
4.14/4.37
31.99
4.22/4.65
3.93/4.20
213.4
72.4
PixelART
✗
✗
3.05B (1.05+2)
36.92
4.32 / 4.91
4.57 / 4.81
36.39
4.26 / 4.87
4.44 / 4.75
3.1 (+0.7)
6.6 (+4.2)
LayerD † [ 54 ]
✓
✓
0.27B
36.07
4.12/3.94
3.49/3.13
37.22
4.06/4.01
3.64/3.37
15.0
3.4
Table 1: System-level comparison . T2I Pre. : whether initialized from pretrained T2I. Pixel PSNR : PSNR between the alpha-composited reconstruction and the input image. VLM : scores (1–5) from GPT-5.5 / Gemini-3.1 Pro (separated by ‘/’) on alpha and visual quality. Efficiency : per-image latency and peak memory on one H200. Parenthesized values for PixelART and CLD report the additional layout detector overhead (#parameters, latency, GPU memory). LayerD † is non-diffusion (segmentation + inpainting), which inflates Composite PSNR but yields the lowest VLM Visual score. Bold = best within the diffusion-based group.
Figure 6: Qualitative comparison with state-of-the-art approaches. QIL: Qwen-Image-Layered.
Table 8
Figure 7: Ablation visualizations. Left: qualitative effect of perceptual losses on background reconstruction. Right top: training loss curves of Muon vs. AdamW under the same recipe. Right bottom: I2L fine-tuning loss curves for T2I-pretrained vs. from-scratch initializations on PixelART-B.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
PixelART-B
PixelART-L
PixelART-H
PixelART (main)
Architecture
# Transformer blocks
12
24
32
32
hidden dim d
768
1024
1280
1280
# attention heads
12
16
16
16
patch size P
16
32
input noise scale
1
2
Appendix
Table 10: Configurations of PixelART variants. PixelART-B is the default for all ablations and is also the small model in the model-size ablation, alongside PixelART-L and PixelART-H. PixelART (main model) is the separately trained system-level model reported in Table 1 ; it shares the H scale but uses a coarser patch size and a three-stage resolution curriculum. Settings that are common across multiple columns are merged. The pixel-space versus latent-space comparison (Table 2 ) uses the same PixelART-B architecture, but the latter operates on the RGBA-VAE latents of Qwen-Image-Layered with a patch size of 2, with the parameter count differing by ∼ 1M due to the smaller patch size.
Figure 8: Web interface used in the user study. Each question shows the original input image, two anonymized layer-decomposition outputs (A/B), the two-dimension ternary choice, and optional fields for failure tags and free-text notes.
Layer metrics
Composite metrics
Prediction
L1 ↓
PSNR ↑
SSIM ↑
L1 ↓
PSNR ↑
SSIM ↑
x -pred
0.0453
32.88
0.8995
0.0086
38.55
0.9727
v -pred
0.0526
30.17
0.8691
0.0114
36.08
0.9441
Appendix
Table 11: x - vs. v -prediction at PixelART-H scale ( d=1280 , no bottleneck). Both runs use identical recipes apart from the prediction target. At this width, v -prediction does converge but still trails x -prediction, indicating that the manifold mismatch is not solely a capacity issue.
Composite metrics
Layer metrics
Decoder
L1 ↓
PSNR ↑
SSIM ↑
L1 ↓
PSNR ↑
SSIM ↑
Qwen-Image-Layered (official)
0.0132
32.52
0.9568
0.0363
31.66
0.9289
Ours (300M ViT)
0.0083
33.69
0.9686
0.0254
37.49
0.9569
Appendix
Table 12: RGBA-VAE decoder reconstruction quality on Design-Multi-Layer-Bench at 256×256 . Each decoder takes the official Qwen-Image-Layered encoder’s latent of a ground-truth RGBA layer and produces a reconstruction, which is compared against the original. Both decoders share the same encoder. Layer L1/PSNR are alpha-pre-multiplied (Appendix D ); layer SSIM is alpha-weighted.
single
two
counting
colors
color
position
overall
obj.
obj.
attr.
PixelART-B (BLIP-3o, T2I)
74.38
24.24
18.12
61.70
12.75
7.50
33.12
Appendix
Table 13: GenEval scores of the BLIP-3o-pretrained PixelART-B (168M params, batch 1024, 250K iters), evaluated before I2L fine-tuning. Overall is the per-task average. The model does not adopt any specialized optimization recipes common in dedicated T2I models, yet still produces non-degenerate outputs across all six task categories at a level consistent with this parameter scale and training budget. This checkpoint serves as the T2I-initialized starting point for the I2L ablation in Table 3 .
Figure 9: Zero-shot decomposition on natural images. Although trained only on multi-layer design templates, PixelART produces coherent foreground/background separation and plausible occluded-region completion on selected natural-image inputs.
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.
Recent advances in pixel-space diffusion models have narrowed the image quality gap with latent-space diffusion, but still converge more slowly and lag behind in final image quality. We argue that a key reason is the lack of an explicit representation prior: unlike latent diffusion, which usually denoises in a compact and structured latent space, pixel diffusion needs to learn denoising-friendly representations and pixel generation simultaneously from raw RGB space. To address this problem, we propose PixelDiT2, an end-to-end pixel-space diffusion model designed to decouple representation learning from pixel generation without introducing an autoencoder or latent reconstruction bottleneck. We propose representation grounding that uses a frozen pretrained vision foundation model to provide explicit per-patch representation guidance throughout denoising, allowing the pixel diffusion transformer to focus more on pixel generation. On ImageNet-256x256, PixelDiT2 achieves an FID of 1.46 after 600 epochs; at 512x512 resolution, PixelDiT2 achieves an FID of 1.48 after 680 epochs. Project page: https://pixeldit.github.io/pixeldit2/
High-quality image editing requires precise modification of target regions while preserving the content of unrelated regions. Existing editing models produce flattened outputs that limit precise local manipulation and layer reuse, while decomposition methods typically optimize layer recovery separately from instruction-driven editing. To this end, we introduce LayerEditor, a unified framework that couples instruction-guided image editing with controllable generation of reusable RGBA layers. At its core, In-Context RGB-Alpha Token Modeling jointly represents the edited canvas and an ordered sequence of RGB-Alpha layers within a single token sequence, enabling semantic edits and explicit control over individual layer content. To retain the pretrained model's editing capability during adaptation to layered outputs, we propose Compositional Edit-Prior Distillation, which transfers its editing prior to both the edited canvas and the image composed from the predicted layers. To support training and evaluation, we construct LayerEditor-100K, a large-scale paired image-editing dataset with ordered RGBA layer annotations, and introduce dedicated benchmarks for layered editing. Extensive experiments demonstrate that LayerEditor maintains strong editing fidelity while generating accurate RGBA layers, supporting both flexible layer-level manipulation and multi-layer decomposition.
Wen Xue, Yong Liu, Xueheng Li +5
South China University of Technology · Tsinghua University · University of Science and Technology of China +1