LayerGrab samples: real image-to-layers results

Six images split into layers by LayerGrab, saved exactly as the model returned them: the original, the filled-in background, every layer as a transparent image, and a JSON manifest with each layer's name, stacking order and position. They are among the samples on the LayerGrab home page, published here so anyone comparing layer-decomposition models, or building something that reads layered results, has real data to look at.

Sample	Image	Layers	Split in
coffee-poster	Coffee shop poster, 1760×2368	10	100 s
headphones-banner	Product banner, 2720×1536	7	66 s
skincare-social-post	Instagram post, 2048×2048	7	80 s
ai-desk-illustration	Isometric desk illustration, 2352×1760	9	100 s
growth-report-slide	Presentation slide, 2560×1600	11	97 s
sportswear-flat-lay	Flat-lay photo, 2416×1616	15	148 s

Layer counts exclude the filled background. index.json lists the same table with the source credit of every sample.

Layout of a sample

samples/<id>/
  sample.json      title, size, seconds, optional prompt, credit, and the layer list
  original.webp    the flat input image
  base.webp        the background with every layer removed and the hidden areas filled in
  layer-01.webp …  one transparent image per layer, numbered by z-index (1 = bottom)

sample.json:

{
  "id": "skincare-social-post",
  "title": "Skincare Instagram post",
  "width": 2048, "height": 2048, "seconds": 80,
  "layers": [
    { "file": "layer-07.webp", "name": "Skincare Serum Bottle", "description": "…", "zIndex": 7, "box": [x1, y1, x2, y2] }
  ]
}

box is in the base image's pixel coordinates; a layer file is the size of its box, not of the whole image. To rebuild the original, draw base.webp, then each layer at its box in ascending zIndex.

What to look at

• Completed elements. In coffee-poster, the cup behind the pumpkin comes back as a whole cup; in ai-desk-illustration the desk under everything is its own whole layer.
• Filled backgrounds. Every base.webp has the elements painted out, not cut out.
• Targeted splits. A sample split with a one-sentence instruction ("only the person on the right") will be added; the tool takes such an instruction on any image.
• Names. Layer names come from the model, in English, as returned.

Licence and credits

Split results and the images made for this demo: CC BY 4.0, attribution "LayerGrab, https://layergrab.com". sportswear-flat-lay/original.webp is a photo by mr lee on Unsplash under the Unsplash License; coffee-poster and headphones-banner were generated with gpt-image-2, ai-desk-illustration and skincare-social-post with Seedream 5.0; growth-report-slide is a mock slide with illustrative figures.

Made with LayerGrab, which splits images like these in the browser or inside Photoshop, Figma, Sketch and Chrome.