The Flat Image Problem

A Picture That Refuses to Stay Still

In December 2025, the Qwen image team published a paper built around an irritation that anyone who has edited generated images knows intimately. You ask to change one thing, and the model changes three other things that were already correct. A face gets subtly rebuilt. A product shifts a few pixels. A shirt changes texture while the background gets a different depth of field. Sometimes the result is excellent, but it is excellent in a slightly different universe from the input.

That problem has not disappeared in August 2026. The current generation of editors is vastly better than the systems people were using a few years ago. GPT Image 2 is OpenAI's current image generation and editing model. Google's current family includes Gemini 3.1 Flash Image, called Nano Banana 2, and Gemini 3 Pro Image, called Nano Banana Pro. Black Forest Labs recommends FLUX.2 for new image generation and editing work. ByteDance has Seedream 5.0 Lite. Alibaba has already moved the Qwen line forward to Qwen-Image-3.0-Pro in limited preview, while its broader image stack also pushes Wan 2.7 for editing. These are not toys, and Stable Diffusion is not the useful reference point for judging them.

But better preservation and guaranteed preservation are different things. A modern editor can be impressively disciplined about leaving a person's face alone. It can be trained on reconstruction tasks, rewarded for consistency, conditioned on several reference images, and given explicit masks or control signals. All of that improves the probability that the untouched parts survive. The Qwen-Image-Layered paper asks whether probability is the wrong place to solve the problem.

The paper's answer is almost embarrassingly concrete. Stop treating the image as one indivisible thing.

A normal raster image is a single rectangular array of pixels. The semantic categories we care about are not represented as independent objects inside that file. The jacket, the face, the lettering, the reflection, and the wall behind the person are all baked into the same final surface. We humans can point at the jacket and say, change only this. The file itself does not contain a jacket object with a clean boundary, a hidden back side, and an independent position. It contains colored pixels that happen to depict a jacket.

That matters because a generative editor generally does not perform the same operation as Photoshop's move tool. It interprets an image plus an instruction and synthesizes an output that should satisfy both. Even when the architecture has strong conditioning on the original image, it is still solving a generation problem. The output is a new image that should look as though only the requested thing changed. That is why identity preservation is a model capability rather than a mathematical property of the file.

Qwen-Image-Layered changes the file representation after the fact. Give it one ordinary red, green, blue image, and it tries to reconstruct the scene as a variable number of red, green, blue, alpha layers. Alpha is the transparency channel. Instead of one flattened picture, you get a stack of transparent sheets. The sheets are meant to be semantically separated enough that a useful object, text element, foreground structure, or background region can live on its own layer.

The important phrase is not simply layer decomposition. It is what the authors call inherent editability. Once the chair is isolated on one layer and the wall is on another, moving the chair no longer asks a generative model to preserve the wall. The wall file is never touched. If the chair layer changes, the other layers can remain byte-for-byte the same. The preservation problem has been moved from model behavior into representation.

That is the first big idea in this paper. A lot of progress in image editing has been about making a model better at obeying a sentence like, change the jacket and nothing else. Qwen-Image-Layered asks whether the stronger solution is to arrange the image so that the jacket and everything else are physically separate before the edit begins.

What A Layer Actually Buys You

There is an easy way to undersell this paper. You could hear layers and think that Qwen has reinvented background removal. It has not. A useful layer decomposition has to solve a nastier problem than segmentation.

Suppose a person stands in front of a patterned wall. A segmentation model can tell you which visible pixels belong to the person and which visible pixels belong to the wall. But if you want to move the person ten centimeters to the right, you need to know what the wall looks like behind the person's old position. Those pixels do not exist in the flattened input. A useful background layer therefore has to contain plausible hidden content. The model is not merely assigning pixels to categories. It is reconstructing complete layers that can survive being separated.

Now add hair, glass, smoke, shadows, reflections, translucent fabric, antialiasing around text, and objects that overlap each other. A binary mask is not enough. Transparency is continuous. A half-transparent pixel may contain contributions from more than one layer. If the decomposition gets that mixture wrong, you see halos, dirty edges, missing shadows, or a foreground object that looks cut out with office scissors.

This is why the paper outputs RGBA layers rather than RGB crops plus masks. Each layer has its own color information and its own alpha matte, and the layers are ordered. If you composite them in sequence using ordinary alpha blending, the stack should reconstruct the original input. The full stack is therefore both a semantic interpretation and an image reconstruction.

There is another subtle benefit. Position becomes external to generation. If an object is a layer, translating it ten pixels, scaling it, hiding it, or recoloring it can sometimes be done with deterministic graphics operations instead of asking an image model to regenerate the scene. This distinction is huge. A generative editor is good at inventing what a red leather jacket should look like. A graphics operation is good at moving an already-correct jacket exactly twelve pixels without deciding that your face could use a new cheekbone.

The paper demonstrates this with resizing and repositioning. Its comparison model is Qwen-Image-Edit-2509, which was a strong Qwen editor at the time the experiment was run. The layered system can perform elementary spatial operations precisely because the object is already separated. Qwen-Image-Edit-2509 has to interpret the desired layout and generate a new coherent image, which can introduce small shifts elsewhere.

But there is a catch that matters. The guarantee begins after decomposition. If the decomposition puts part of the chair shadow into the wall layer, or splits one logical object across several layers, the system has not magically solved scene understanding. It has made a prediction about how the flattened image should be taken apart. That prediction can be wrong.

And there is no single true decomposition for many images. Should a person's shadow be part of the person layer, a separate shadow layer, or baked into the ground? Should the letters of one sign be one layer or several? Should a pair of glasses belong with a face or with accessories? Photoshop files answer these questions according to whoever built the document, not according to physics. The paper acknowledges this ambiguity in its evaluation by allowing adjacent predicted layers to be merged when matching a reference decomposition.

So the representation gives a hard preservation property for layers you do not touch, but obtaining the representation is still a learned inference problem. That distinction is the key to understanding both why the idea is powerful and why it is not a universal editing solution.

The 2026 Competition Is Smarter Than The Paper's Villain

The paper's introduction frames global generative editing as fundamentally vulnerable because it resamples the whole image. That criticism is structurally correct, but the practical comparison has moved quickly since the paper was submitted on December 17, 2025.

OpenAI now describes GPT Image 2 as its state-of-the-art image generation model for both generation and editing, with high-fidelity image inputs. Google has pushed native image generation deeper into Gemini, with Nano Banana 2 designed around generation, conversational editing, multiple references, consistency, world knowledge, and up to 4K output. FLUX.2 supports multi-reference editing with several model tiers and up to 4 megapixel output. Seedream 5.0 Lite explicitly adds more reasoning and current-information support. Alibaba's Qwen-Image-3.0-Pro preview now targets dense layouts, fine text, photorealistic detail, and combined generation and editing.

None of those facts invalidate the paper. They sharpen the question. If whole-image editors become good enough that unwanted drift is rare, do we still need explicit layers?

For some tasks, maybe not. If you want to turn a summer photograph into winter, change the camera angle, invent a different hairstyle, or combine three reference people into one scene, a generative editor is doing exactly the kind of global reasoning you want. A neat layer stack may even be an awkward intermediate representation because the desired transformation affects lighting, reflections, geometry, and context across the whole image.

For other tasks, probabilistic preservation remains the wrong contract. Consider a product photograph where the logo must be exactly unchanged. Consider a newspaper advertisement where one price needs to move but every other line must remain fixed. Consider a character sheet where you have finally got the face right and never want a model to reinterpret it again. In those cases, there is a qualitative difference between a model being very good at preserving pixels and a workflow in which those pixels are not sent through the edit operation at all.

This creates a useful way to divide modern image work. Some edits are semantic synthesis. They require a model to understand what the scene means and invent coherent new pixels. Other edits are structural manipulation. They require exact control over content that already exists. Today's strongest systems increasingly try to do both through one conversational model. Qwen-Image-Layered instead argues that we can make the representation itself carry some of the structural burden.

That is why Stable Diffusion deserves only a historical footnote here. The interesting comparison in 2026 is not between layers and an old latent diffusion pipeline with weak text understanding. It is between explicit structured images and current multimodal editors that are astonishingly good at acting as if they had structure internally, even when their output is still a flat raster.

The next episode goes inside Qwen's mechanism. Not at code level, but one floor below the interface: how a model that normally generates a single image is altered so it can produce an unknown number of transparent layers, why the variational autoencoder has to change, and why Qwen adds a third positional dimension for layer number.