How AI layer splitting works
LayersApart sends your image to ByteDance’s Seedream 5.0 model with layer decomposition switched on. The model looks at the whole picture, decides which elements stand on their own, and returns one image per element plus a background with those elements removed and the gaps painted in.
What comes back
- A background (always the bottom layer) with the separated elements removed and filled in.
- Up to 16 transparent PNG layers, each with a short English name, a description and a box that says where it sits on the background.
- The stacking order, so a cup on a table sits above the table.
Each layer file is usually larger than the box it belongs in. LayersApart scales every layer into its box before showing the preview or building the PSD, and the ZIP keeps the original full-resolution files with a layers.json file that lists each box.
Why the stack is not identical to your image
A layer is not a cutout of visible pixels. The model completes each element, including parts that were hidden behind something else, and repaints the background behind everything it lifts. That is what lets you move an element and find a whole scene underneath, but it also means some details are redrawn.
In a published test on one sample image (LaoZhang AI, 2 October 2026), restacking the layers changed 30% of pixels with Flash and no instructions, 12.6% with Pro, and 4% with Flash when only two elements were named. Treat those numbers as one example, not a guarantee. Use Show original in the viewer to compare before you download.
Getting better results
- Name what you need. “Separate only the bottle and the headline” returns two layers plus the background. In the same test, naming two elements cut the time from about 95 to 35 seconds and the image count from 14 to 3.
- Use Pro for busy scenes or out-of-focus areas you want left alone.
- Start from a clean, large image. Heavy JPEG compression and tiny images give rough edges.
- Text stays raster. Lettering comes back as its own image layer. Retype it in your editor if you need different words.
Limits
- Input: one PNG, JPEG or WebP (converted to PNG), at least 512 × 512 pixels in total, aspect ratio no longer than 16:1. LayersApart resizes very large images to a maximum of 2560 pixels on the long side before upload.
- Output size: 1K or 2K. 1K returns a background around one megapixel; 2K is sharper and takes longer.
- A split takes roughly 30 seconds to 2 minutes. It fails as a whole if the model cannot finish; failed splits are not charged.
- Results are kept for 23 hours.
Sources
- BytePlus ModelArk, Seedream 5.0 Pro tutorial (layer decomposition parameters), as quoted in the LaoZhang AI test below. We could not read the BytePlus page directly when writing this.
- LaoZhang AI, “Seedream 5.0 Pro Layer Decomposition API Tutorial”, published 2 October 2026, read 8 October 2026.