Research map / Image generation / Text-to-image diffusion

PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Layout-to-image control · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains PixelART, a from-scratch pixel-space rectified-flow Transformer with x-prediction and terminal-boosted timestep sampling, for region-conditioned image-to-layer RGBA decomposition, matching or beating VAE-based T2I-pretrained systems at lower cost.

Problem

Image-to-layer decomposition (splitting a flat image into editable ordered RGBA layers) is currently handled by large pretrained text-to-image diffusion models with RGBA autoencoders, which are heavy and can bottleneck fine details, transparency, and boundaries.

Approach

Model I2L directly in pixel space with a single-stream ViT-style rectified-flow Transformer trained from scratch on regional RGBA patches, predicting clean pixels (x-prediction with v-loss), and oversample high-noise timesteps (terminal-boosted sampling) because layer assignment is decided early in denoising.

Key results

• PixelART-H achieves Composite PSNR 36.92 on Design-Multi-Layer-Bench vs 32.46 for Qwen-Image-Layered and 29.79 for CLD, with ~6x fewer parameters than QIL.
• PixelART-H runs 19x faster than CLD and 69x faster than QIL on 1024x1024 inputs, using 8-11x less peak GPU memory (3.1 s, 6.6 GB).
• Pixel-space diffusion beats latent-VAE diffusion by +5.57 dB Composite PSNR (34.36 vs 28.79) in matched PixelART-B ablation; terminal-boosted U(0,0.2) with p=0.1 gives Layer PSNR 29.51 vs 28.66 baseline.

Surprising findings

• T2I pretraining on ~36M pairs gives marginal or slightly negative final gains (Composite PSNR -0.39 dB) for I2L despite lower initial loss.
• v-prediction fails to converge at PixelART-B scale (Composite PSNR 9.05 vs 34.36 for x-prediction).

Limitations

• Training data is an internal, non-public design corpus, limiting reproducibility.
• Requires region bounding boxes as input; full automation depends on an external layout detector (DAD, internal).

Experimental setup

Datasets: internal 4M multi-layer design templates BLIP-3o corpus (ablation only) Design-Multi-Layer-Bench LICA

Benchmarks: Design-Multi-Layer-Bench LICA

Metrics: Layer L1 Layer PSNR Layer SSIM Composite L1 Composite PSNR Composite SSIM VLM Alpha score VLM Visual score alpha-IoU owner accuracy

Baselines: Qwen-Image-Layered CLD LayerD Magic Layers

Builds on: JiT SD3 Muon Qwen-Image-Layered ART

Closest papers by meaning

• PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
• PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.87)
• Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.87)
• PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis (ICLR 2024 · similarity 0.87)
• Decomposition of Graphic Design with Unified Multimodal Model (ICML 2025 · similarity 0.87)
• MULAN: A Multi Layer Annotated Dataset for Controllable Text-to-Image Generation (CVPR 2024 · similarity 0.87)