Splitting a JPG Into Layers Is Becoming a Normal Design Step

A quiet shift is happening in the tools designers reach for. For most of the last year, the interesting action in AI image work was generation: making something that did not exist. The tools that are multiplying now start from an image that already exists and take it apart.

The shift is easy to miss because it does not produce dramatic demo reels. Nobody shares a screenshot of a well-separated layer stack. But it is the change that decides whether generated images are usable in professional work, because a design asset's value depends less on how it looks and more on whether it can be edited later.

LayerGrab appeared on October 2. It converts a flat PNG, JPG or WebP into up to sixteen independent transparent layers, with descriptive English names, and reconstructs the background behind extracted elements through inpainting. It exports individual PNGs, a ZIP, or a layered PSD for Photoshop, with plugins for Photoshop, Figma and Sketch and a Chrome extension for grabbing images from a browser. There is no .psd export from the browser tool itself; the layers come as numbered PNGs that you stack in an editor.

The processing takes 60 to 100 seconds. It runs on segmentation and generative models including Qwen-Image-Layered, the model Alibaba open-sourced on October 2 under a permissively licensed release. Input files are capped at 20MB, and a Chrome extension lets a designer right-click an image on any site and send it straight to the separation panel.

Why layer decomposition is the useful half

Generation solved a problem designers had been working around for years, and created a new one. An AI-generated image looks like a flat photograph of a design, not a working file. You cannot move the product to the other side of the banner, because it is welded into the pixels. You cannot replace the headline text, because there is no text object, only the appearance of text.

That is precisely the gap layer decomposition fills. The point here is recovering the structure a flat image hides, so the image can be edited the way a design file is edited, rather than producing something new.

Alibaba's framing of Qwen-Image-Layered describes the problem as "flat pixel coupling." The model uses an RGBA-VAE encoder and layer-level 3D positional encoding to identify spatial levels and occlusion relationships, then decomposes the image into independent editable layers. The company's claim is near-zero drift in precise edits, which addresses the consistency problem that shows up whenever you change one part of an image and watch other parts shift.

Ant Group's inclusionAI released a related pair on the same day, Ming-Image-0.1-Design for generating complete visual designs and Ming-Image-0.1-Design-Layer for decomposing flattened designs into transparent layers, both 6B parameters, both requiring an 80GB GPU for default deployment.

What creators actually do with it

The tool descriptions read as a list of small, unglamorous jobs. Split a finished ad into layers and move the product to the other side for a new format. Separate an illustration into layers to animate a parallax effect. Pull the product and background of a photo apart so they can animate on separate slides. Extract elements from a poster to change the text or reuse the graphics in a new project.

None of those are tasks that made anyone's demo reel. They are the tasks that consume a designer's afternoon, and they were previously not automatable at all because no tool could recover structure from a rendered image. A designer facing a flat JPG had exactly one option: redraw it by hand, layer by layer, guessing at what sat behind each element.

The reason this matters more than it sounds is that most assets a designer receives are flat. A client sends a JPG of last year's poster and asks for a version in a new format. A brand team has the final export but lost the source file. A campaign asset exists only as a screenshot. Layer decomposition turns those dead ends into starting points, which is a different kind of value than generation provides.

There are honest limitations. LayerGrab does not preserve cast shadows. Very fine text may show slight color variations when redrawn by the model, and text is extracted as an image layer rather than editable type. As with any segmentation system, the split can vary slightly between runs on the same input.

Kenerate's Image Layer tool takes a similar approach with a different emphasis, producing two to eight layers with the user choosing the count, or an automatic count that often lands between six and twelve on its Pro tier. It also exports PNGs rather than a PSD, and the documentation is explicit about that limit, which is the kind of honesty a tool in this category benefits from. Designers who expect a working PSD file and receive a folder of PNGs waste an afternoon discovering the difference.

The variation between tools is instructive. One optimizes for a fixed layer count a designer controls; another lets the model decide. Both route around the same constraint, which is that layer segmentation is a judgment call and different tools make it differently.

The bigger direction

Look at what shipped this month as a group and a pattern shows up. Alibaba open-sourced a layer model. Ant Group released a design-and-decompose pair. A commercial tool wrapped the capability into plugins for the editors designers already use. The capability went from a research result to a product feature in about two weeks.

That speed is the interesting part. When a model is open-sourced with a permissive license, the window between "this exists in a paper" and "this is a button in Figma" collapses. Designers do not adopt architectures; they adopt buttons. The model that wins is rarely the one with the best benchmark score. It is the one that shows up inside the software already open on the designer's second monitor.

The practical effect is that AI images stop being terminal artifacts. An image generated today can be disassembled tomorrow, rearranged on Thursday, and shipped in a format the client's team can still edit next quarter. That flexibility is what makes generated assets usable in professional workflows, and it is why layer decomposition has moved from novelty to a standard step.

The next question is whether layer decomposition ends up inside the generation model or beside it. If a model produces a layered output directly, the separate decomposition step disappears. Until then, tools that recover structure from flat images are the bridge between a generator's output and a designer's requirements, and the bridge is being built fast enough that most design teams will use it without ever deciding to adopt it.