Seedream 5.5 Layer Split: Turn a Flat AI Image Into Editable Layers, and What It Really Costs
ByteDance's new image model will take a finished poster apart into separate transparent pieces. For designers, that matters more than any quality score. Here is how to use it, what to pay, and where the seams show.
Every AI image so far has shipped as a single sheet of pixels. The product, the shadow under it, the headline and the background are fused together, and the only way to move the headline is to ask for the whole picture again and hope nothing else changes.
ByteDance's Seedream 5.5 offers something different. Hand it a finished image and it hands back a stack of parts: a base picture plus separate transparent PNG files for the objects it recognized, each with a label, a position and a place in the stacking order. Runware, one of the services selling it, says the split can return "up to 99 transparent PNG layers." That sounds like a Photoshop file from a model that has never seen one.
It is close to that. It is also not quite that, and the gap is where your time and money go.
Why a layer split matters more than a better picture
If you design for clients, you know the most expensive sentence in the job: "Love it, can we just change the headline?"
With a normal AI image, the honest answer is no, not without risk. Regenerating to fix one word can shift the product, recolor the background or redraw a face the client already approved. So designers end up painting over AI output by hand, or rebuilding it from scratch in a layout tool, and the AI saved them less time than it promised.
A layer split changes the shape of that request. The headline is its own layer. The product shot is its own layer. You can swap one, nudge another and leave the rest untouched, with ordinary tools you already own.
Seedream 5.5 pushes the same idea from two other directions. ByteDance's official page leads with "Multi-turn editing, consistent every time," and says edits keep the untouched parts of the picture "intact at the pixel level." It also promises "native support for transparent and semi-transparent backgrounds," so a product or a glow can come out on its own with no cut-out step. Put together, the pitch is simple: stop rerolling, start editing.
What it is and where to get it
Seedream 5.5 is ByteDance's image generation and editing model. Its official page carries no release date. Several of the services that sell it, including Segmind, date it to November 9, 2026, and Runware's changelog added it on November 12.
There is no download. You use it through a hosted service. The ones with public pages:
- BytePlus, ByteDance's own cloud business, advertises it as "starting at $0.02/image" and links a playground from ByteDance's page.
- Segmind lists about $0.021 per image at 1K size and about $0.037 at 2K, plus a small fee for each reference image.
- Runware lists two cents per image up to 1.5K and three and a half cents at 2K, plus $0.008 per reference image.
- fal sells the layer split as its own tool, called layerize, at $0.03 per returned image at 1K or 1.5K and $0.0405 at 2K, plus $0.012 per input image.
AlphaSignal reports that it is also available through Dreamina and CapCut, ByteDance's consumer apps, which is the easiest route if you never want to touch an API. I could not confirm that from ByteDance directly, so treat it as a lead.
No service I read states the license in plain words on the model page. fal shows a "Commercial use" badge. For paid client work, read the terms of whichever service you pick.
How the layer split works, in plain terms
You give it one finished image. Runware allows only one reference image when splitting, so this is a "take this apart" job, not a "combine these" job.
The model looks at the picture, decides which things are separate objects, and returns:
- A base image with those objects removed and the space behind them painted in.
- A transparent PNG for each object, with a name, a short description and a box that says where it sits.
- Optionally, a text reading of any words in the design. On Runware it returns the text, its position, its color and the closest-matching typeface (which can come back as "unknown"). Artistic and hand-lettered type gets no reading.
That third part is easy to overlook, and it may be the most useful piece for designers. If the model tells you the headline is set in something close to a particular typeface, you can replace the AI's lettering with real type in that face and never fight its spelling again.
You then stack the layers in Photoshop, Figma, Affinity or anything that reads PNGs, using the positions it gave you. You get loose PNGs, not a ready-made PSD, so plan on a few minutes of assembly or a small script.
What the hands-on test found
The numbers in launch coverage ("100 rounds of editing," "99 layers," near-perfect text) come from ByteDance as relayed by resellers. The only detailed public test I found is Segmind's, and it is worth reading in full because it is honest about both sides.
The good news first. Segmind ran 12 editing rounds on one image and found the background held up: almost no pixels drifted far from where they started. Transparent product cut-outs came back with clean edges. Soft effects like a glow came out partly see-through, the way they should.
Now the seams.
The split was smaller than the headline. Segmind's layer tests returned between 2 and 8 layers. Its example request asked for at most 10 images, and Runware allows up to 99 layers plus the base, so this does not prove the model cannot do more. It does mean nobody I found has shown you 99.
The restacked file is close, not identical. When Segmind placed the layers back on the base at their stated positions, the result differed from the original by 9.4 percent of its pixels. On a social graphic, you will not notice. On a packshot the client signed off on, you might.
The background behind each object is invented. The model has to guess what was behind the product. Move the product and you reveal that guess.
Glass failed. Asked twice for a see-through glass, the model returned it solid both times.
Text reading missed things. It skipped the main title, "EMBERWOOD," in one test, and returned a button as a picture layer with its words left behind in the base image.
Small pieces came back blown up. Small layers returned at about one megapixel, so you will scale them back down.
Long edit chains darken. Across those 12 rounds, the whole frame got darker by about half a level per round. One round is invisible. Twenty would not be.
The cost math nobody puts in the headline
Two cents an image is the generation price. A layer split bills you for every image it returns, base included.
Segmind's own nine-image split (a base plus eight layers at 1.5K) cost $0.199. Runware's documentation prices its own split examples at $0.208, without saying how many layers they returned. That is still cheap next to an hour of masking, but it is ten times the number in the headline, and it adds up if you split every draft.
The sensible habit: generate and choose cheaply, then split only the image the client actually picked.
Put this into practice this week
Here is a one-hour test that tells you whether Seedream 5.5 belongs in your workflow.
1. Pick a real job you already finished. A social graphic or a simple poster with a product, a background and a headline. You want an image where you already know what "right" looks like.
2. Run the split once. Use fal's layerize page, Segmind or the BytePlus playground. Turn on text reading if the design has type. Ask for a modest number of layers, around six to eight, so you are not paying for slivers.
3. Rebuild it. Drop the base and the layers into your design tool at the positions returned. Put the original on top and set it to Difference blend mode. Anything that lights up is where the rebuild departs from the original. That is your honest picture of the 9.4 percent.
4. Make the edit a client would ask for. Move the product left. Replace the headline with real type in the suggested typeface. Recolor the background layer. Time yourself.
5. Compare with your old way. If the edit took five minutes and the old way took forty, Seedream earns a place in your kit. If you spent the time fixing the base image behind the product, it does not, at least not for that kind of picture.
6. Write down what broke. Glass, fine hair, thin type, reflections. Those are the cases to keep away from the split, or to hand-finish.
Where it falls short
Seedream 5.5 is hosted only. There are no model files to run on your own machine, so every split depends on a service staying up and keeping its price.
The headline numbers are claims. ByteDance's page itself does not state "100 rounds" or "99 layers"; those come from resellers quoting ByteDance. Segmind's careful test covered 12 rounds and up to 8 layers.
The split invents what it cannot see. That is fine for flat graphics and risky for photography where the hidden area matters.
Licensing is per service. Check it before anything goes into paid work.
And it does not replace a designer's eye. It hands you parts. Deciding which parts to trust, and which to redraw, is still your job.
The real shift
The interesting thing about Seedream 5.5 is not that the pictures look better. It is that the model finally treats a picture the way designers always have: as a stack of things you can pick up and move. If that idea holds, the reroll stops being the default way to revise AI work, and the edit becomes the default again.
Try it on one finished job before you trust it on a live one. Your own Difference layer will tell you more than any launch chart.