Seedream 5.5 API

Seedream 5.5 is ByteDance's new image generation and editing model, announced by the ByteDance Seed team on 9 November 2026 and rolling out this week across the company's apps and cloud platforms. It is the first point release after the Seedream 5.0 family, and it is aimed squarely at production work: images that have to survive a long round of client changes, assets that have to drop straight into a layout, and posters with small print that has to be readable.

Three upgrades carry the release. The model can now generate images with a real alpha channel, so a product, a sticker or an icon comes out with a transparent background instead of needing a cutout. It holds an image steady across long editing sessions, with ByteDance demonstrating 100 consecutive edits on one portrait while the untouched parts stayed identical at the pixel level. And it renders light, skin, materials and dense small text more faithfully than Seedream 5 Pro, the model it follows.

On Unicorn API, Seedream 5.5 ships as two endpoints: text to image, with the new transparent-background switch, and layer decomposition, which takes a finished image apart into a base plus as many as 99 transparent layers and can read the text in them.

Where Seedream 5.5 comes from

Seedream is built by ByteDance Seed, the foundation-model group that also makes the Seedance video models and the models behind the Doubao assistant. The image line has moved at a steady clip. Seedream 4.0 arrived in September 2025 and merged generation and editing into one model. Seedream 4.5 followed in December with better consistency and typography. Seedream 5 Lite opened the 5.0 generation in February 2026, Seedream 5.0 Pro in July introduced point-and-box editing and layer separation for designers, and Seedream 5.0 Flash in September brought those tools to a faster tier.

Seedream 5.5 is the next step on top of 5.0 Pro, and ByteDance measures it directly against that model. The early-access API id carries the version stamp 261028. It is rolling out in Jimeng (Dreamina), Doubao, the Doubao Work canvas and Volcano Engine Ark, and third-party hosts listed it within days. ByteDance has also said a Pro version of Seedream 5.5 will follow, without a date.

Like every Seedream release, 5.5 is closed. There are no downloadable weights and no licence for self-hosting, and ByteDance has not published a technical report, a parameter count or an architecture description. What is known about the model comes from the Seed team's announcement, the API documentation and early hands-on tests.

What changed from Seedream 5.0 Pro

Stable multi-round editing. Long edit chains are where image models usually fall apart: each pass shifts colours a little, softens faces and nudges the background, until by round ten the picture is visibly not the one you started with. ByteDance's headline claim for 5.5 is that unedited regions stay pixel-identical with no colour shift, and its demonstration put one portrait through 100 rounds of changes to clothing and accessories, with small objects such as rings and glasses still editable cleanly at the end. Independent testing is more modest but encouraging. One host ran twelve rounds on a product shot and found no background pixel moved by more than 32 levels, while single-object edits touched only about 0.5 to 1 percent of the frame.

Native RGBA output. Seedream 5.5 can produce an alpha channel directly, with both fully transparent and semi-transparent pixels. ByteDance highlights hair, fur, glass and lace edges, the details that a background remover usually mangles.

Realism. Light, materials and skin texture are the second focus. In ByteDance's human evaluations on its MagicArena platform, 5.5 scored +55.62 percent on a good/same/bad preference measure over the previous version for realism in image-to-image work.

Dense small text. The Seed team reports a severe text corruption rate of 0.4 percent and a minor one of 1.7 percent across Chinese, English and dense small lettering, and says complex character shapes are reproduced more accurately.

Usable output by scenario. ByteDance tested six production scenarios: e-commerce, marketing, film and television, gaming, office documents and fashion design. Quality scores and usable-image rates rose in all six, and in film and television and in office work the usable rate more than doubled compared with 5.0 Pro.

Against other labs. ByteDance says 5.5 leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure in both generation and editing, with a substantial lead in text-to-image realism. These are the company's own numbers; it has not published the prompt sets or evaluator counts, and no independent leaderboard has confirmed them yet.

Speed. Chinese launch coverage puts 5.5 at nearly 40 percent faster than 5.0 Pro. One host measured an average of 24 seconds per image, ranging from 18 to 39 seconds depending on the request.

Transparent backgrounds without a cutout step

The transparency feature is easy to underrate until you look at how assets are made today. A typical pipeline generates a product on a plain backdrop, runs a background remover, then cleans the fringes by hand. Seedream 5.5 skips all of that. On Unicorn API you set background to transparent with output_format set to png, and the model returns a PNG whose alpha channel was generated with the image, including soft partial transparency on edges and shadows.

The other settings are opaque, which forces a solid background, and auto, which lets the model decide from the prompt. JPEG output cannot carry transparency, so a transparent request needs PNG.

This matters most for e-commerce catalogues, stickers and emoji packs, game sprites and UI illustration, and for any layout where a generated element sits on top of a designed background. Testers have found one consistent gap: asking for a transparent glass or a glass of liquid tends to return an opaque object, so truly see-through materials still need work after the fact.

Layer decomposition with text recognition

The second endpoint runs in the opposite direction. You send one finished image, a JPEG, PNG or WEBP up to 10 MB, and the model returns a base image plus separated elements as transparent PNGs: the subject, headline text, decorative shapes, the background. Each layer comes with a name, a description, its stacking order and a bounding box, so your own code can reassemble or rearrange the picture.

Compared with the 16-layer ceiling of Seedream 5.0 Pro and Flash, 5.5 goes much further. The max_images parameter accepts 2 to 100 outputs including the base image, with a default of 30, which makes busy infographics, menus and slide-style images practical to unpick.

The new text_recognition switch is the other upgrade. When it is on, layers the model recognises as text also carry the text itself, line boxes and style runs such as font, colour and weight. That turns a flat poster into something close to an editable design file: translate the headline, keep the font and colour, put it back where it was.

Without a prompt the model picks the main elements on its own. With a prompt of up to 2,000 characters you can name what to separate, or mark regions with <bbox>x1 y1 x2 y2</bbox> tags. One early test found that reassembling layers from their boxes left about 9 percent of pixels slightly off the original, so treat the layers as a starting point for editing rather than a lossless split.

Specs at a glance

Where Seedream 5.5 shines

Ready-to-place assets. Product shots, icons, stickers and characters with a clean alpha channel come straight from a prompt. For marketplace listings and ad templates, that removes a whole step.

Small print. Recipe cards, menus, infographics, packaging mock-ups and slide-like images with several sizes of type are where the text improvements show. Early testers found Spanish and Japanese copy reproduced exactly, accents and all.

Photographic realism. Skin, fabric, metal and window light are rendered with more texture and less of the waxy finish that marks many generated portraits.

Long revision cycles. The editing stability is aimed at the way commercial work actually goes: a dozen small changes from a client, each of which must not disturb what was already approved.

Turning flat images into editable files. Decomposition with text recognition is useful for localisation pipelines, template builders and anyone who needs to rework an image they no longer have the source file for.

Limits worth knowing

The headline numbers are ByteDance's. The text corruption figures, the six-scenario results and the comparison with OpenAI and Google models all come from the company's own evaluations, without published prompt sets or confidence intervals. The 100-round demonstration is a stress test, not a typical session.

Long chains are stable, not perfect. In one test the whole frame darkened by about half a level per round, small on its own but noticeable over very long chains. An edit that moved a subject also changed the composition, turning one shoe into a pair.

Text can reshape the scene. The model sometimes redesigns a surface to make room for requested text, and one tester saw two letter errors in a small Hindi subhead. ByteDance's text figures cover Chinese and English; check other scripts carefully, and check anything regulated, such as packaging labels, line by line.

2K is the ceiling. There is no 3K or 4K tier. Large print work needs an upscaler or Seedream 5 Lite.

See-through materials. Transparent output handles hair and soft edges well but tends to return glass and liquids as opaque.

No open weights. You cannot fine-tune Seedream 5.5 or run it on your own hardware.

How to prompt Seedream 5.5

Seedream follows structured prompts well: subject first, then setting, light, style, and any exact text in quotes. For transparent assets, describe the object alone and say what should be soft or semi-transparent; leave out backgrounds, tables and floors.

A transparent asset (set background to transparent, output_format to png):

A glazed ceramic teapot in deep cobalt blue with a bamboo handle, three-quarter view, a thin wisp of steam rising from the spout, soft studio light, isolated object, no background

Dense small text:

A printed recipe card on off-white paper, title "Lemon Ricotta Pancakes" in bold serif at the top, below it a two-column ingredient list in small sans-serif type with eight items and quantities, a numbered method of five short steps, a small line drawing of a lemon in the corner, flat print style

A realistic portrait:

Close-up portrait of an elderly fisherman mending a net on a harbour wall at dawn, weathered skin, wool cap, cold blue light with a warm glow from the low sun, shallow depth of field, 85mm photograph

A guided decomposition (layer decomposition endpoint, text_recognition on):

Separate the headline, the label text, the product photo and the background pattern

Tips: aspect_ratio is required, so set it explicitly. Explore at 1K and re-run the keepers at 2K. Use PNG for anything you will edit or layer later, and keep max_images near the number of elements you actually need so the response stays manageable.

Seedream 5.5 versus other models in the catalog

• Seedream 5 Pro is the predecessor, with point-and-box editing and 16-layer decomposition. Seedream 5.5 improves realism and small text, adds native transparency and goes far beyond 16 layers.
• Seedream 5.0 Flash is the fast tier of the 5.0 family and, on Unicorn, the Seedream with an image-to-image endpoint for coordinate edits today.
• GPT Image 2.5 Sunburst is OpenAI's high-end option with strong instruction following and conversational editing; ByteDance's own tests put 5.5 ahead on text structure and realism.
• Nano Banana 2.1 is Google's general-purpose generator and editor and a natural second opinion on any prompt.
• FLUX.3 [klein] is the pick when open weights matter.
• Recraft Remove Background cuts out an existing image; Seedream 5.5 makes the transparent asset in the first place.

The short version: if your output goes into a layout as an asset, or carries a lot of small text, Seedream 5.5 is the strongest ByteDance option available today.

Getting started

Open the Seedream 5.5 playground to try prompts in the browser, or switch to layer decomposition to split an image. For code, create a key on the API keys page and follow the docs: send a job to /v1/jobs with the model id seedream-5-5-text-to-image, a prompt, an aspect_ratio and optionally size and background, then poll for the result. The API tab on this page has complete examples.

Verdict

Seedream 5.5 is a production-minded upgrade rather than a flashy one. Native transparency removes a step from every asset pipeline, the text and realism gains show up where commercial images most often fail, and layer decomposition with text recognition turns flat renders into material a designer can rework. The benchmarks are ByteDance's own, output stops at 2K and glass still comes back solid, but for teams that generate images to put into real layouts, it is the most practical Seedream yet.