MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale

TL;DR

MRT is a 20B-parameter masked region diffusion model for multi-layer transparent image generation and editing, trained on >10M multilingual design samples with diverse aspect ratios. It unifies text-to-layers, image-to-layers (decomposition), and layers-to-layers (editing) under a single masked region diffusion framework, where selective token masking selects which task to run. Two main technical contributions: (a) the unified masked-region task formulation, and (b) an overflow-aware canvas layer that handles boundary inconsistencies and lets layers extend beyond the visible canvas with semi-transparent backgrounds. Diffusion distillation brings inference down to 8 steps for real-time generation. The headline result is that MRT substantially outperforms commercial systems and the concurrent Qwen-Image-Layered model on image-to-layers user studies, while running 10-100× faster and cutting activation GPU memory by 50-90% at I2L inference. No weights released at filing time.

Key claims

A single masked region diffusion framework can unify text-to-layers, image-to-layers, and layers-to-layers tasks at the 20B parameter scale, selecting tasks via selective token masking rather than separate heads [§Abstract].
An overflow-aware canvas layer mechanism handles layers whose content extends beyond visible canvas boundaries, supporting semi-transparent background synthesis and complete editable layers — addressed as a first-class problem rather than cropped away [§Abstract].
Diffusion distillation brings multi-layer generation to 8 steps with “minimal quality degradation,” enabling real-time inference at the 20B scale [§Abstract].
MRT “significantly outperforms the concurrent Qwen-Image-Layered model” in image-to-layers quality on user-study evaluations [§Abstract].
At I2L inference MRT achieves 10-100× faster wall-clock and 50-90% lower activation GPU memory than Qwen-Image-Layered [§Abstract].
Training scale: 10M+ multilingual design samples spanning diverse aspect ratios and textual prompts [§Abstract].

Method

MRT is a 20B-parameter MMDiT-class masked region diffusion model. The unifying primitive is selective token masking: at training and inference, different patterns of masked vs. visible tokens across the (text, source image, target layers) sequence select the active task. Text-to-layers masks all layer tokens; image-to-layers reveals a source RGB and masks the layer tokens; layers-to-layers reveals a subset of layers and masks the layers being edited. A single shared backbone learns all three jointly.

The second contribution is the overflow-aware canvas layer: a designated extra layer that represents content extending beyond the visible canvas, together with the boundary-handling so that semi-transparent background regions and off-canvas extensions are reconstructed coherently. This addresses a long-standing artifact in layered decomposers where layers get cropped at canvas boundaries and downstream edits (resize, reposition) expose the missing content.

Real-time inference is achieved via diffusion distillation down to 8 steps. The abstract describes the resulting model as supporting “real-time, multi-layer generation.”

The Slack post notes weights are not yet released (paper only); only paper-side abstract details are available at filing time. The arXiv abstract is the only retrievable source — section anchors above reflect that.

Results

Image-to-layers user study: MRT beats Qwen-Image-Layered (Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition) significantly, and also beats various unnamed commercial systems across all three tasks (T2L, I2L, L2L) [§Abstract].
Inference: 10-100× faster than Qwen-Image-Layered at I2L; 50-90% lower activation GPU memory during I2L inference [§Abstract].
8-step distilled inference with “minimal quality degradation” vs. the full-step teacher [§Abstract].

Quantitative tables (FID, alpha IoU, RGB L1, etc.) and bench-by-bench numbers are in the paper PDF and not in the abstract retrieved at filing.

Why it’s interesting

This is the second wiki entry in two months to scale image-to-layers decomposition to the ~20B-parameter Qwen-Image regime — and the first to explicitly target the same head-to-head as Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition. Two architectural choices distinguish MRT from the Qwen-Image-Layered line: (a) instead of a layer-index axis in 3D RoPE (the route both ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation and Qwen-Image-Layered take), MRT uses selective token masking over a shared MMDiT sequence to encode which layers are inputs vs. outputs vs. edits — the same primitive that unifies inpainting and editing in many MMDiT editors, repurposed for layer-task selection; (b) MRT introduces an explicit overflow-aware canvas layer rather than letting clipped-at-canvas content remain a known artifact. The 10-100× inference speedup over Qwen-Image-Layered at I2L is the kind of claim that, if it holds up, makes the layered-decomposition primitive cheap enough to drop into design tools at production scale — though no weights are out, so independent verification is pending. The 8-step distillation result also makes MRT a natural data point for Diffusion Distillation at 20B scale in the layered-image regime, which prior cluster entries have not covered.