The Photoshop Files Behind the Model

Human Layers Are A Mess

A model cannot learn to take images apart without examples of images that have already been taken apart. That sounds obvious, but it creates a data problem. The internet contains an absurd number of finished JPEG and PNG images. It contains far fewer high-quality examples where the final picture and its meaningful transparent component layers are both available.

Qwen's answer is wonderfully practical. Find Photoshop documents.

A Photoshop document can contain the composite image people eventually export plus the editable layer stack used to build it. That makes it unusually valuable supervision for this task. Instead of inventing synthetic foregrounds and backgrounds, the researchers can observe how real designers separated text, photographs, effects, shapes, and other elements while making actual visual work.

But a Photoshop document is not a clean scientific description of a scene. It is a work file. Anyone who has opened somebody else's layered design knows what that means. There can be hidden layers, duplicated experiments, adjustment fragments, abandoned objects, layer groups, masks, effects, dozens of tiny pieces, and names like Layer 37 copy 2 final final. A document may contain hundreds of layers even though a human viewer would describe the picture as having a handful of meaningful elements.

The researchers therefore build a data-cleaning system around the PSD files. They extract the layers with the open-source psd-tools library. They remove layers with anomalous content, giving blurred faces as one example. They discard layers that do not contribute to the final composite. And when layers occupy different parts of the image without spatial overlap, they can merge them to reduce the total number of output layers.

That last choice is technically useful and philosophically revealing. The model is not trying to recover the sacred original layer stack exactly as a designer created it. It is trying to recover a useful decomposition whose layers can reconstruct the image and support editing. If two elements never overlap, putting them on one sheet may preserve the ability to edit the composition while reducing the computational burden.

The authors also use Qwen2.5-VL to generate text descriptions for the composite images. Those captions support the text-conditioned stages of training. So the data has three related forms: the final flattened picture, the transparent layer stack, and a language description of the content.

There is an important limitation here. Photoshop layer structure is human production structure, not ground-truth physics. A designer may bake a shadow into an object layer. Another may put the shadow on its own layer. A photographer may flatten an entire photograph and place only typography above it. An illustrator may separate every decorative stroke. The dataset teaches the model a useful human editing convention, but that convention is not a unique decomposition of reality.

That helps explain why the paper later allows predicted layers to be merged when it scores them. There can be several valid ways to partition the same visible picture.

Teaching It In Three Acts

The training itself is staged because the authors are asking a pretrained image generator to change several habits at once. If they simply replaced the VAE, added a layer coordinate, and immediately demanded correct decomposition, the model would have to learn too many unfamiliar things simultaneously.

Stage 1 is text to RGBA. The model already knows how to generate ordinary RGB images from text. Now it is trained jointly on text-to-RGB and text-to-RGBA tasks using the new four-channel VAE. This is the adaptation stage where transparency becomes part of the model's visual language without discarding its ability to make normal images.

Stage 2 expands from one RGBA image to several. The authors introduce text-to-multiple-RGBA generation. The model learns to predict both a final composite and the transparent layers that make it up. This gives the new layer coordinate system something concrete to do. It also forces information to move between the complete picture and the component sheets.

They call the result of this branch Qwen-Image-Layered-T2L, meaning text to layers.

Stage 3 adds the task the paper is really about: image to multiple RGBA layers. Instead of starting only from text, the system now receives the flattened RGB image and learns to decompose it. This branch is Qwen-Image-Layered-I2L, image to layers.

The reported training schedule is substantial. The three stages run for 500,000, 400,000, and 400,000 optimization steps. The maximum number of layers is set to 20. The optimizer is Adam with a learning rate of one times ten to the minus five. Those numbers matter less than the shape of the curriculum. The model first learns the new pixel format, then the new output structure, then the inverse problem of recovering that structure from a flattened input.

That curriculum is one of the strongest lessons in the paper because it is transferable beyond layers. When a capable pretrained model has to learn a new representation and a new task, there is often value in separating those changes. First teach it the new alphabet. Then teach it the new sentence structure. Only then ask it to solve the hard problem in that language.

The authors test this claim with an ablation study. An ablation means deliberately removing components to see what breaks. Their table is cumulative rather than a full combinatorial experiment, so it should not be read as a perfectly isolated measurement of every component. Still, the direction is very clear.

With the new layer coordinate, RGBA-VAE, and multi-stage training all removed, the system performs badly. With no layer merging allowed during evaluation, its alpha soft intersection-over-union score is about 0.37. Add Layer3D RoPE and it rises to about 0.58. Add the shared RGBA-VAE and it reaches about 0.65. Add the full multi-stage training and it jumps to about 0.87.

The RGB reconstruction error shows the same broad pattern. Lower is better. The stripped-down version reports about 0.28. The full system reports about 0.059. The point is not that one magic component wins. The paper's engineering story is that the representation, positional structure, and curriculum reinforce one another.

What The Benchmarks Actually Prove

This is where a good paper deserves a mildly suspicious reader.

For quantitative decomposition evaluation, the authors use the Crello dataset and follow a protocol introduced by LayerD. Because the predicted stack and reference stack can contain different numbers of layers, the evaluator aligns the sequences with an order-aware form of dynamic time warping. It also permits adjacent layers to be merged. That is a sensible accommodation for the fact that decomposition is ambiguous.

They score two things. RGB L1 measures color error, weighted by the reference alpha. Lower is better. Alpha soft intersection-over-union measures how well the predicted transparency matches the reference transparency. Higher is better.

With zero layer merges allowed, Qwen-Image-Layered reports an RGB L1 of 0.0594 and alpha soft intersection-over-union of 0.8705. LayerD reports 0.0709 and 0.7520. If the evaluator is allowed to merge up to 5 adjacent layers, Qwen reaches 0.0363 and 0.9160, while LayerD reaches 0.0396 and 0.8650. On this benchmark, under this protocol, Qwen is clearly better.

There is a detail that should stay attached to that result. The authors say their PSD-derived training distribution differs substantially from Crello, so they fine-tune Qwen-Image-Layered on the Crello training set before evaluating on the Crello test set. That is a normal experimental choice, but it means the number is not a pure zero-shot demonstration that the PSD-trained model automatically dominates on a foreign layer convention.

The RGBA-VAE gets its own reconstruction test on AIM-500. There, Qwen's VAE reports the best values across the paper's four metrics compared with the listed transparency-capable baselines. One of those tables includes SDXL-based and FLUX-based components. In August 2026, that should not be misread as a ranking of complete modern image generators. It is a component-level transparency reconstruction comparison from the paper's experimental setup.

The paper's strongest headline is about editing consistency, but the editing evidence is mostly qualitative. The authors compare examples against Qwen-Image-Edit-2509 and show cases where the ordinary editor shifts pixels or struggles with moving and resizing elements, while a decomposed layer can be translated or scaled directly. The mechanism is convincing: if an untouched layer is never regenerated, it cannot drift. But the paper does not provide a broad 2026-style head-to-head editing benchmark against today's strongest proprietary and open systems.

That distinction matters. The paper demonstrates that layered representation creates a different preservation contract. It demonstrates strong decomposition scores on Crello. It demonstrates that its RGBA-VAE reconstructs transparent images well. It does not demonstrate that every real photograph will decompose into perfect professionally useful layers, or that the complete workflow dominates every current editor on every edit.

There is also a conceptual trap in the phrase exactly unchanged. Once the layer stack exists, untouched layers can indeed stay exactly unchanged. But the act of creating that stack is generative. Hidden background content has to be invented. Transparent boundaries have to be estimated. Semantic grouping has to be chosen. If the decomposition is wrong, exact preservation can faithfully preserve the wrong structure.

This is not a fatal criticism. It tells us where the intelligence has moved. A conventional editor spends intelligence on every edit, trying to regenerate the whole scene while preserving most of it. Qwen-Image-Layered spends a large amount of intelligence up front to infer a structured document. After that, some edits become ordinary graphics operations with deterministic behavior.

That shift is especially interesting for repeated editing. A one-off transformation may not justify decomposition. A design that will be revised twenty times might.

And research after the Qwen paper suggests the core idea did not end with this one system. In early 2026, LASAGNA explored jointly generating photorealistic backgrounds and transparent foregrounds with effects such as shadows and reflections. Referring Layer Decomposition introduced prompt-conditioned extraction of specific complete RGBA layers and a dataset with more than a million image-layer-prompt triplets. The research problem is already branching from generic decomposition toward controllable, user-directed layer creation.

The final episode moves from the paper to August 2026. Qwen itself has advanced two generations in product naming, OpenAI, Google, Black Forest Labs, and ByteDance have stronger unified editors, and newer layer research has appeared. The interesting question is no longer whether layers can beat an old editor. It is where explicit structure still gives you something that a very smart flat-image model cannot promise.