Seedream 5.0: Flash, Lite, Pro and API Code (2026)

Seedream 5.0 is ByteDance’s current family of image generation models. It covers text-to-image, reference-based editing and, new to the line, splitting a finished image into separate layers. If you are trying to work out which version is which, whether it beats Google’s Nano Banana models, and how to call it from code, this page covers all three and gives you working requests you can paste into a terminal.

What is Seedream 5.0?

Seedream is ByteDance’s text-to-image line. A text-to-image model takes a plain-language prompt and returns a picture that matches it. Most current models, Seedream included, also accept reference images, so you can edit or restyle a picture you already have. Under the hood these are diffusion models: they start from random noise and remove it step by step, guided by your prompt, until an image appears.

Seedream 4.0 made the line known for high-resolution output and strong editing (see our Seedream 4.0 pricing and editing guide). The 5.0 generation keeps that base and adds what ByteDance describes as multi-step reasoning. Before it renders, the model works out the layout of the scene, where objects sit relative to each other and which domain details matter. In practice you notice it in three places:

• Complex scenes. Prompts with several subjects and explicit positions (“the mug left of the laptop, the plant behind both”) are followed more faithfully than older models managed.
• Text-heavy images. Infographics, posters and labelled diagrams are a stated focus, especially for the Pro variant.
• Targeted edits. You can ask for a change in one region of an image and expect the rest to stay put most of the time.

When was Seedream 5.0 released?

The family arrived in stages. The Lite variant came out in February 2026, and a Pro variant aimed at professional generation and editing followed later in the year. The Flash variants, the fast tier, became available on Unicorn API on 29 September 2026.

Some descriptions of Seedream 5.0 mention real-time web search during generation, so the model can pull in current information. Treat that as a feature of specific apps built on the model. The Flash endpoints covered below take a prompt and images, and none of their input fields enables search.

Seedream 5.0 Lite, Pro and Flash: which variant is which?

The naming is confusing because different sites expose different tiers. Here is the short version:

• Lite is the lighter, earlier release. It is a good fit for general text-to-image and image-to-image work where speed and cost matter more than maximum detail.
• Pro is the professional tier. It is described with native 4K output, deeper reasoning, precise layer editing and strong text rendering for infographics and spatially targeted revisions.
• Flash is the fast tier. On Unicorn API it comes as three separate models, one per task.

The three Flash models are:

1. Seedream 5.0 Flash Text To Image: prompt in, image out. Inputs are prompt, size, aspect_ratio, nsfw_checker and output_format.
2. Seedream 5.0 Flash Image to Image: a prompt plus one or more reference images in image_urls (an array). Use it for edits, restyles and keeping a product or character consistent.
3. Seedream 5.0 Flash Layer Decomposition: takes a single image in image_url (a string, not an array) and a prompt, and splits the image into layers you can edit separately.

All three accept sizes of 1K, 1.5K and 2K, and the price is the same at every size.

A note for anyone searching for the Seedream 5.0 Lite price: Lite pricing depends on the platform that hosts it, and Unicorn API does not list Lite. If cost is why you were looking at Lite, Flash is the closest option here, and it is cheap (see the pricing section below).

What Seedream 5.0 does well, and where it struggles

A model is only useful if you know where it breaks. After running a fair number of prompts, here is the honest picture.

Strengths

• Layout adherence. Positional instructions (“three bottles in a row, tallest on the right”) come out right far more often than with older diffusion models.
• Edit consistency. With the image-to-image model, changing a background or a colour usually leaves the subject’s shape, logo and lighting intact.
• Graphics with text. Short headlines, labels and simple infographic layouts render legibly. Keep text short and put it in quotes in the prompt.
• Speed and price. The Flash tier is built to be fast, and its flat per-image price makes 2K output cheap.

Weaknesses

• Portrait framing. The most common complaint is that subjects look great but the model chooses its own crop. Ask for a full-body shot and you may get a medium shot. Ask for a centred subject and it may sit off to one side.
• Over-polished look. Default output leans toward glossy, high-contrast commercial imagery. If you want something candid or documentary, say so explicitly (“natural light, slight grain, unretouched”).
• Long text blocks. A headline works. A paragraph of body copy usually does not. Put long copy in your layout tool instead.

The framing problem has practical fixes. Set the image shape with the aspect_ratio field instead of describing it. Put framing first in the prompt and use photographic terms (“head and shoulders portrait”, “full body, feet visible, subject centred”). If it still drifts, pass a reference image with the composition you want to the image-to-image model, which anchors the layout much harder than words do.

How to call the Seedream 5.0 API

Every image model on Unicorn API uses the same pattern. You create a job, poll it until it finishes, then download the result. That means the code you write for Seedream 5.0 works unchanged for the other image models in the model catalog.

1. Create an API key on the keys page and export it as UNICORN_API_KEY.
2. Send POST /v1/jobs with the model id and an input object.
3. Poll GET /v1/jobs/{id} until status is "succeeded".
4. Read the image URL from outputs[0].url.

curl

# 1. Create a text-to-image job
curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedream-5-flash-text-to-image",
    "input": {
      "prompt": "Minimalist poster for a coffee brand, the words \"MORNING RITUAL\" in bold sans-serif at the top, a ceramic cup centred below, warm beige background",
      "size": "2K",
      "aspect_ratio": "1:1"
    }
  }'

# 2. Poll with the id from the response
curl https://api.unicornapi.net/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $UNICORN_API_KEY"

Python

This helper creates a job, waits for it and returns every output URL. You will reuse it in the worked example below.

import os
import time
import requests

API = "https://api.unicornapi.net"
HEADERS = {"Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}"}


def run_job(model, input_, timeout=180):
    r = requests.post(f"{API}/v1/jobs", headers=HEADERS,
                      json={"model": model, "input": input_})
    r.raise_for_status()
    job = r.json()

    deadline = time.time() + timeout
    while job["status"] != "succeeded":
        if job["status"] == "failed":
            raise RuntimeError(f"Job failed: {job}")
        if time.time() > deadline:
            raise TimeoutError(f"Job {job['id']} still {job['status']}")
        time.sleep(3)
        r = requests.get(f"{API}/v1/jobs/{job['id']}", headers=HEADERS)
        r.raise_for_status()
        job = r.json()

    return [o["url"] for o in job["outputs"]]


if __name__ == "__main__":
    urls = run_job("seedream-5-flash-text-to-image", {
        "prompt": "Head and shoulders portrait, subject centred, a chef in a "
                  "white jacket in a busy kitchen, natural window light",
        "size": "2K",
        "aspect_ratio": "1:1",
    })
    print(urls[0])

The full list of job statuses and error shapes is in the API documentation. If you work inside Cursor or Claude Code, you can also call these models from your agent through MCP; see using the API from agents. To try prompts before writing code, the image playground runs the same models in the browser.

Worked example: edit a product photo, then split it into layers

Here is a realistic job. You have a plain photo of a skincare bottle on a white table. You want a seasonal campaign version, and your designer wants the result in layers so they can move the headline around.

Step 1: restyle with image to image

edited = run_job("seedream-5-flash-image-to-image", {
    "prompt": "Keep the bottle, label and logo exactly as they are. "
              "Place it on a mossy stone in a misty autumn forest, "
              "soft morning light, add the headline \"FALL GLOW\" "
              "in clean white serif type in the top third",
    "image_urls": ["https://your-cdn.example/bottle.png"],
    "aspect_ratio": "1:1",
    "size": "2K",
})[0]
print("Edited:", edited)

The first sentence of that prompt is doing the real work. Telling the model what must not change (“keep the bottle, label and logo exactly as they are”) before describing the change cuts down on unwanted redesigns of the product.

Step 2: decompose the result into layers

layers = run_job("seedream-5-flash-layer-decomposition", {
    "prompt": "Separate the headline text, the product bottle and the background",
    "image_url": edited,
    "size": "2K",
})
for i, url in enumerate(layers):
    print(f"Output {i}:", url)

Notice the field change: image_urls (array) for image to image, image_url (string) for layer decomposition. The script loops over all outputs instead of assuming there is only one. Check the model page for exactly how the layers come back.

What it cost

Two calls at the Flash price come to 8 credits once rounding is applied (explained below), which is $0.08 on a monthly plan. A designer doing the same composite by hand would spend far longer, and if the first edit misses you can rerun step 1 a few times for well under a dollar.

Seedream 5.0 vs Nano Banana Pro and other image models

The comparison most people search for is Seedream 5.0 vs Nano Banana Pro, Google’s top-tier Gemini image model. Both target the same work: reasoning-heavy scenes, readable text and consistent edits. Seedream’s advantages are speed and price at the Flash tier, plus a dedicated layer decomposition endpoint. Google’s models are a strong choice when you lean on many reference images or need higher resolutions. For a deeper look at Google’s side, see how to use Nano Banana.

This table compares the Flash tier with the alternatives you can call through the same Unicorn API key:

Model	Maker	Resolution options	Reference images	Credits per image
Seedream 5.0 Flash (all three)	ByteDance	1K, 1.5K, 2K	Via image_urls (image to image)	3.24 at any size
Nano Banana 2.1	Google	1K, 2K, 4K	Via image_input	4 / 6 / 9
Nano Banana 2 Lite	Google	1K	Up to 10	4
GPT Image 1.5	OpenAI	Low / medium / high quality	Up to 4	3 / 4 / 22
GPT Image 2.5 Flare	OpenAI	1K, 2K, 4K	Up to 16	6 / 10 / 16

How to read this:

• For cheap 2K images, Seedream Flash is the obvious pick. At 2K it costs less than Nano Banana 2.1 (6 credits) or GPT Image 2.5 Flare (10 credits).
• For 4K in one call, use Nano Banana 2.1 or GPT Image 2.5 Flare. Flash stops at 2K.
• For transparent backgrounds, GPT Image 1.5 and GPT Image 2.5 Flare expose a background field. Seedream’s route is layer decomposition.
• For many references at once, such as a character sheet plus outfits plus a location, GPT Image 2.5 Flare (up to 16) and Nano Banana 2 Lite (up to 10) are built for it.

Spec tables only take you so far. The way to decide is to take ten prompts from your real workload, run them through two or three of these models, and look at the results side by side. Because they share one API and one request shape, that comparison is a change of the model string. For a wider survey of editing models, see our best AI image editor roundup and the photorealistic image generator comparison.

How much does Seedream 5.0 cost?

On Unicorn API, each Seedream 5.0 Flash model is listed at 3.24 credits per output, the same at 1K, 1.5K and 2K. Prices round up to whole credits, so a single image bills as 4 credits. Since size does not change the price, generate at 2K unless you have a reason not to.

What a credit costs depends on how you buy it (full details on the pricing page):

• Monthly plans give credits at $0.01 each: Starter $25, Creator $50, Studio $100, Pro $200 per month. At 4 credits per image, Starter covers about 625 Flash images a month.
• Top-ups cost $0.015 per credit and never expire, which works out to $0.06 per Flash image. That suits occasional use.

You never have to guess. POST /v1/models/price returns the exact price of a request before you run it, so you can show costs in your own app or guard a batch job against overspending:

curl -X POST https://api.unicornapi.net/v1/models/price \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedream-5-flash-image-to-image",
    "input": {
      "prompt": "Same bottle on a marble counter",
      "image_urls": ["https://your-cdn.example/bottle.png"],
      "size": "2K"
    }
  }'

Common mistakes with Seedream 5.0

Most failed or disappointing runs come down to a handful of avoidable errors:

1. Describing the shape instead of setting it. “Wide landscape banner” in the prompt is a suggestion. aspect_ratio is an instruction. Use the field.
2. Mixing up image fields. Image to image wants image_urls as an array, even for one image. Layer decomposition wants image_url as a single string. Sending the wrong one is the most common error on first integration.
3. Using URLs the model cannot reach. Reference images must be publicly fetchable. Links behind a login, localhost addresses or signed URLs that expire before the job starts will fail.
4. Reading outputs too early. outputs is only populated once status is "succeeded". Poll, do not assume.
5. Expecting 4K from Flash. The Flash endpoints top out at 2K. Plan for an upscaling step, or choose a model with a 4K option.
6. Forgetting what must stay the same. In edits, say what to keep before what to change. Otherwise the model may “improve” your logo.
7. Stuffing prompts. Twenty adjectives dilute each other. One clear subject, one setting, one lighting note and the framing will beat a keyword list.
8. Using one model for everything. Seedream 5.0 is a strong default for fast, cheap 2K work and layered edits, but it is not the best tool for every job. Keep a second model in your pipeline for 4K or heavy multi-reference tasks. With one API key, switching costs you nothing.

If you want to start right away, create a key, paste the Python helper above into a file and run the text-to-image call. Your first Seedream 5.0 image will cost 4 credits, and you will know within a minute whether it suits your work.