XinYu.
ナビゲーション
開発者
MCPCLI
言語
← ニュースに戻る
実測2026/11/12

Seedream 5.5 Layer Split, Tested: 49 Layers from 5 Images — What Works and What Doesn't

Seedream 5.5 Layer Split, Tested: 49 Layers from 5 Images — What Works and What Doesn't

XinYu AI Testing Team · Tested on Nov 12, 2026

Seedream 5.5's layer split turns one image into a base plate plus a set of transparent layers. We ran it once on each of 5 very different images, put every layer back in place, and compared the result with the original.

The short version: posters and flat illustrations come out close to ready to use. Fog, cast shadows and glow in photoreal images don't separate cleanly yet.

Left: the original poster. Right: some of the 10 layers it was split into

How we tested

  • All 5 source images were generated with Seedream 5.5 at 2K: a portrait by a window, a text poster, a product shot, a foggy courtyard, and a flat illustration.
  • Layer split at 2K, layer cap 20 (the cap includes the base), no prompt — the model decided what to split.
  • One run per image. Running the same image again may give a different result.

The results in one table

ImageLayers (excl. base)TimeWhat workedWhat didn't
Window portrait4~1 minHair edges; the street behind her was filled inOnly 4 layers; the window frame changed color in the base
Text poster10~1 minOne layer per line of text, clean edgesText is pixels, not editable type
Product shot5~1 minBottle, stand and leaves all separatedThe splash was repainted bluish; shadows lost
Foggy courtyard16~2 minPine, lanterns, figure split in fine detailAll the fog stayed in the base; reassembled, it looks like paper cut-outs
Flat illustration14~2 minFurniture, cat, wall art each on its own layerLamp light on the wall and the chair's shadow are gone

The layer cap is a ceiling, not a target. With the same cap of 20, the portrait gave 4 layers and the fog scene 16. The image decides. You're charged for the cap up front and anything the model doesn't return is refunded automatically, so a higher cap doesn't cost more — your balance just needs to cover the cap up front.

The base plate fills in what was hidden

This is the most useful part. The base isn't a cut-out with holes: whatever the foreground covered gets redrawn. The street and bikes behind the woman, the starry sky under the poster text, the empty room behind the furniture are all complete.

Top: portrait original and base. Bottom: illustration original and base

The base is a new painting, not the original. In the portrait, the window frame is dark wood in the original (it became its own layer), but the frame in the base is white. If you only want the base as a clean background, check details like this.

Each layer has a name and a position, but you can't just stack the files

Every layer comes with a name, a one-line description, and its position in the original. Names and descriptions are in English, e.g. "Off-white event time text" or "Sleeping orange tabby cat".

Each layer file is upscaled on its own. The "NOV 21 · 8 PM" line on the poster covers about 660×255 px of the base, but its layer file is 3288×1275 px. Drop the PNGs straight into an editor and every size and position is wrong. Each layer has to be scaled down and placed using its position box.

We checked all 49 layers across the 5 images: each layer's aspect ratio matches its position box within 1.2%, so placing by the box lines everything up. Downloading from the XinYu canvas does this for you (see the end).

What splits well: graphic design and illustration

Top: hair, pine needles and small vertical text. Bottom left: the floor-lamp layer picked up a strip of window frame. Bottom right: a distant building in the fog came out as a blur

  • Text poster: title, subtitle, date, venue and button each got a layer, and both small vertical text blocks were split out too. Edges are clean. Every layer is still an image, so changing the words means retyping them.
  • Flat illustration: 14 layers — furniture, cat, wall art, plants — and every object's shape and position lines up when reassembled (lighting is another story, see below).
  • Portrait hair: the curls are cut out strand by strand, with no white fringe.

What doesn't: fog, shadows, light

Left: original. Middle: base. Right: all layers reassembled

1. Fog and smoke don't get their own layer. In the courtyard, all the fog stayed in the base; the pine, steps and rocks came out crisp and fog-free. Reassembled, things that were half hidden in mist now sit sharply on top of it, like paper cut-outs. The temple deep in the fog came out as a semi-transparent smudge.

2. Cast shadows and lighting get lost. The shadows of the stand and leaves on the tabletop, and the warm lamp light on the wall and the chair's shadow in the illustration, belong to neither an object layer nor the base, so they disappear. The lamp layer did keep a halo around the shade.

Left: original. Middle: reassembled. Right: the splash layer

3. Layer content can be repainted. The splash is clear water in the original, but its layer came back filled with pale blue in the middle, which is obvious once reassembled. Leaf angles and plant outlines shift slightly too.

4. Occasional bleed between layers. The illustration's floor-lamp layer picked up a strip of the window frame on its left.

Comparing each reassembled image pixel by pixel with the original, all 5 are much closer than the base alone, but none is indistinguishable:

ImageBase only vs originalAll layers vs original
Window portrait12.3 dB24.6 dB
Text poster10.3 dB23.5 dB
Product shot15.6 dB21.4 dB
Foggy courtyard15.1 dB23.8 dB
Flat illustration9.9 dB19.2 dB

(PSNR; higher means closer to the original. The illustration's objects all line up yet it scores lowest — most likely because the wall light and shadows are gone.)

Left: illustration original. Right: 14 layers reassembled

When it's worth using

  • Good fit: posters, graphic layouts, flat illustrations, product shots on clean backgrounds. Move, recolor or delete elements with little cleanup.
  • Use it as "cut out the subject + fill the background": in photoreal images, the isolated subject and the filled-in base are both useful on their own. Just don't expect the stacked layers to equal the original.
  • Poor fit: photoreal scenes dominated by fog, smoke, water, glass or strong shadows. You'll be repainting shadows and atmosphere afterwards.

Getting a usable PSD in XinYu

  1. Select an image on the canvas, click Split layers in the image toolbar, and pick the quality and layer cap.
  2. When it's done, click Show layers to show or hide layers one by one.
  3. Click Download all layers and choose:
    • Layered PSD: every layer already scaled and placed, with names and order matching the canvas; layers hidden on the canvas are hidden in the PSD too. The fog scene's 17-layer PSD (base included) is 2731×1536 px, about 33 MB.
    • PNG layers (.zip): one transparent PNG per layer at the base's size, so they line up in any editor.

The PSD canvas follows the base, and the base can differ slightly in size from the original (fog scene: original 2816×1584, base 2731×1536). Keep that in mind if you need to register it against the original.

From the command line or MCP, xinyu generate layers and xinyu_split_layers return each layer's name and position box (normalized 0–1000).

FAQ

Can Seedream 5.5 layer split export a PSD directly? The model itself returns a set of individually upscaled PNGs plus each layer's position, not a PSD. On the XinYu canvas, Download all layers gives you a PSD with everything already in place.

What layer cap should I use? We used 20 on all 5 images and got 4 to 16 layers. The cap doesn't force extra layers; you're charged for the cap up front and the unused part is refunded, so a higher cap doesn't cost more, as long as your balance covers the cap up front.

Is the split text editable? No. Text layers are images on a transparent background. The CLI has an extra text-recognition option, which we didn't test here.

What we didn't test

  • Only 5 images, one run each, all generated by Seedream 5.5 itself — no camera photos.
  • No prompt. Whether naming which elements to split (in words or with position boxes) helps with scenes like the fog one is untested.
  • Text recognition was off, and we only tested 2K.

共有

この記事で検証したモデルSeedream 画像生成を XinYu AI で使う →利用できるバージョン、得意なこと、よくある質問はモデルページにまとめています。