The 8 Best AI Image Layer Splitters in 2026, Compared

The same six questions were put to every layer splitter. They are the questions that decide whether a split is usable, not whether it looks impressive in a demo.

1. How many layers, and how sensible. Does a poster with ten elements come back as ten layers, or as “foreground” and “background”?
2. Are the elements whole. When one thing sits in front of another, does the thing behind come back complete, or with a hole where the front object was?
3. Is the background repaired. After every element is lifted off, is what is left a clean picture, or a picture full of cut-outs?
4. Can you ask for one thing. If you only need the price tag, can you say so, or do you pay for and wade through the whole split?
5. Where the layers can go. Copy to the clipboard, PNG files, a layered PSD, or straight into Photoshop, Figma or Sketch, or do they stay inside the tool?
6. What it costs. Per layer, per image, per month, or in GPU time.

What was actually run. LayerGrab was run on seven images: a coffee-shop poster, a headphones banner, an Instagram post, an isometric illustration, a presentation slide, a flat-lay photo and a two-person photo. All seven results are saved unedited and are the samples on the home page. The Qwen-Image-Layered public demo was tried three times with the coffee poster. The other six tools require an account and, in most cases, payment, so they are described from their own product and help pages, each linked with the date it was read. Where a page is silent the table says so. If you run the same images through one of them, the layer names and timings on our side are the baseline to compare against.

LayerGrab is an AI tool that splits one image into up to 16 separate transparent layers, one per element, over a background that is filled in behind them, or pulls out just the object you name. It runs on layergrab.com and, with the same account and credits, as Photoshop, Figma and Sketch plugins and a Chrome side-panel extension.

What we got. The coffee poster (880×1184 px) came back as 10 layers in 100 seconds: the headline, the “20% OFF” badge, four maple leaves, the cup, the drink, the whipped cream and the event line, over a filled background. The flat-lay photo of sportswear gave 15 layers in 148 seconds, one per item down to the wristband. The Instagram post gave 7, the illustration 9, the slide 11. Every layer had an English name we could search for, such as “Yellow +42% Year on Year Badge”.

Whole elements. In the poster, the pumpkin graphic sits on the cup. Split apart, the pumpkin is its own layer and the cup layer is a complete plain cup, not a cup with a pumpkin-shaped hole. In the illustration, the desk that everything stands on is one whole layer. This is the difference between a layer stack you can rearrange and one you can only hide.

Asking for one thing. Type what you want before splitting. “Separate the woman on the right” on the two-person photo returned exactly one layer plus the filled background, in 46 seconds, and cost two layers instead of a whole split. Details are on the separate one object page.

Where the layers go. Copy one layer and paste it into Figma, Photoshop or Canva; download it as PNG; download every layer as a ZIP; or export a layered PSD. In the plugins, the split arrives as a named group directly above the image, aligned to it.

Price. Credits are charged per layer returned, the filled background included: 10 credits a layer on the basic plan ($12 a month for 1,200 credits), 8 on premium, or a one-time pack from $20. New accounts in supported regions get 30 free credits. If a split returns more than the balance covers, nothing is charged and the result is shown locked until it is paid for, so a low balance never pays for half a result.

Where it falls short. Text is pixels, not type. Shadows an object cast on the background are not carried with it. Two runs on the same image can split it slightly differently. And it is a hosted service: the image is sent to LayerGrab's servers, so it is the wrong choice if the picture must never leave your machine.

Magic Layers is Canva's built-in feature that “transform[s] any flat image into fully editable design layers” so you can “select, separate, and edit individual elements”. It launched in public beta in March 2026. You reach it from the Canva home page or from Edit in the editor, wait while it processes, then move, delete, resize, apply effects to or replace the elements.

What the help page says. It works with single-page PNG and JPG files. It does not support PDFs, vector graphics, multi-page documents or licensed images, and lists photo-realistic imagery under “results may vary”: it “works best with graphic designs, illustrations, and stylized visuals”. Images over 5,000 px on the longest edge are not supported. It uses Canva's AI credits, with a monthly limit that depends on your plan. The page does not mention exporting the layers as files or as a PSD.

Where it fits. If your whole workflow is in Canva, this is the shortest path: no download, no upload, the elements land in the design you are already editing. If the layers have to leave Canva, or you work in Photoshop or Figma, it is the wrong tool, which is why we keep a separate Canva Magic Layers alternative page. For photos, Canva's Magic Grab (a different feature) lifts one selected object off a photo and keeps the background.

Sources: Canva help: Turn images into editable layers with Magic Layers; PetaPixel, March 2026. Read 2026-09-29.

Codia's Magic Layers is a browser image layer splitter that promises to “instantly separate backgrounds, objects, text, products, and visual elements into independent layers”. Upload a PNG, JPG or WebP (the page also names screenshots, posters, mockups and AI-generated images), and Codia opens its Studio, places the image on a canvas and “converts the image to layers automatically”. From there you can “select, move, replace, hide, or regenerate individual layers while the rest stays intact”.

What the page says, and does not. It says Magic Layers can be tried free from the page, which makes it one of the easiest ways to see a split on your own image. It names use cases: replacing a hero product, removing background objects, swapping a person, editing decorations, making campaign variations. It does not state a layer limit, a file-size limit, a price for continued use, or what export formats leave Studio. The result lives in Codia's canvas; if you need PNG or PSD files, check what Studio exports before you rely on it.

Where it fits. A first free test, and workflows that continue inside Codia's own editor. There is no way to ask for one element before the split; you select it afterwards.

Source: codia.ai/magic-layers. Read 2026-09-29.

Not to be confused with Canva's or Codia's feature of the same name, magiclayers.net is a stand-alone site that “separates one flat PNG or JPEG into editable layers” and gives “each one back as a named transparent PNG”. Two controls on its upload form set it apart from most of this list: an “Images to generate” slider, where you choose 2 to 15 images (one base image plus layers), and an optional “Describe what to separate” box, with examples such as “Separate all people in the image”.

Limits and price. PNG or JPEG only, up to 10 MB (larger files are downscaled), at least 512×512 px, one image per job. Every layer comes back “named, described, and positioned”. It costs 2 credits per output image, the base image included; plans are $29 a month for 400 credits, $79 for 1,200 and $149 for 2,400, and the page mentions no free tier. It says plainly that the result is not a recovered Photoshop or Figma document: fonts, vectors and effects are not restored.

Where it fits. When you want to decide the layer count yourself and describe what to separate, and PNG files are enough. At $29 for 400 credits, a full 15-image split is 30 credits, so about $2.20; a two-image split is about 15 cents.

Source: magiclayers.net. Read 2026-09-29.

imagetolayers.com is a browser layer splitter that takes a JPG, PNG or WebP and promises to “separate subjects, objects, characters, and backgrounds, reconstruct plausible hidden areas behind overlaps, and keep working with the result as editable PSD or PNG layers”. Output is transparent PNG layers, a ZIP, or a PSD. It says most images take 60 to 120 seconds.

What the page says, and does not. The layer structure “depends on what is actually present”; there is no way to say which element you want. It is honest about its limits: fine detail, heavy blur, extreme overlap and ambiguous boundaries may need editing, and “reconstructed areas are AI-generated estimates, not recovered original pixels”, which is true of every tool on this page. Sign-in is required. A Starter trial gives 20 credits once; Creator is $10 a month (12,000 credits a year) and Studio $30 a month. The page does not state how many credits one split uses.

Where it fits. PSD export at a low monthly price, for people who split whole images and do not need to pick one element.

Source: imagetolayers.com. Read 2026-09-29.

The one open model on this list. Qwen-Image-Layered (Qwen team, weights released December 19, 2025, Apache 2.0, 20B parameters) decomposes an image into a number of RGBA layers you choose, and any layer can be decomposed again. There is an official demo on Hugging Face Spaces, and the layers can be downloaded as a ZIP or exported to PPTX.

What we tried. We uploaded the coffee poster to the demo, set Layers to 10 (the maximum; the default is 4) and clicked Decompose. Three attempts, including one at the default 20 inference steps instead of 50, all stopped with the same message: “The requested GPU duration (360s) is larger than the maximum allowed.” The demo asks for six minutes of GPU per run, which is more than an anonymous visitor is allowed on the free tier, so in practice it needs a Hugging Face login. We did not sign in, so we have no output to show; the interface itself is in the screenshot above.

What the documentation says. The prompt “is intended to describe the overall content of the input image” and “is not designed to control the semantic content of individual layers”, so you cannot ask it for the price tag. You control the count instead: 2 to 10 layers in the demo. Running it yourself needs a CUDA GPU and the diffusers pipeline; the README recommends the 640 resolution bucket for this version. There is more in our plain guide to Qwen-Image-Layered.

Where it fits. Research, pipelines, and anyone whose images must stay on their own hardware. Once it is running, each image is free.

Sources: demo Space; model card; README. Read 2026-09-29.

LlamaGen's AI layer splitter is aimed at illustration: it separates “background, foreground, subject, text, objects, comic characters, anime art, props, or product graphics”. Upload, drop, paste or give a URL. A free Quick Mode produces a PNG layer sheet; a paid PRO Mode “uses an upgraded model for cleaner separation” and unlocks the individual transparent PNGs, ZIP and PSD.

What the page says. Its best-results guidance is specific: one clear cartoon, anime, game, mascot, product or illustrated subject, full silhouette visible, simple background, clean outlines. It warns that “very crowded scenes, hidden body parts, overlapping subjects, sketch marks, heavy shadows, cropped limbs, or tiny accessories may still need source cleanup or manual finishing”, and calls itself “not a shortcut for every crowded photo, sketch sheet, occluded pose, or dense layout”. Credit prices are not given on the page.

Where it fits. Single-character artwork and game or anime assets, especially if you already use LlamaGen. Dense posters and photos are outside what it recommends for itself.

Source: llamagen.ai AI Layer Splitter. Read 2026-09-29.

Image Breakdown is a community-made layer splitter app in Krea's node editor (the page credits its author) that uses Krea's models to “split any image into editable layers with AI”. The workflow on the page: drop a JPG or PNG, choose transparent layers or a layered file as the target, let the segmentation run, then “adjust layer boundaries and masks for cleaner edges” and export PNGs or a layered format.

What the page says, and does not. Its pitch is mask control: cutouts you can refine before exporting. It does not state a layer count, whether hidden areas are filled, or a price; running it needs a Krea sign-in and Krea's plans apply. Because it is one node graph among many, it can change or disappear without notice in a way a product feature would not.

Where it fits. People already on Krea who want to refine masks by hand, and collage or layout work where a clean cut-out matters more than a repaired background.

Source: krea.ai Image Breakdown. Read 2026-09-29.