Seedream 5.0 Flash style blending: a five-step brand board test plan

To test multi-reference style blending, fix one prompt, vary only the reference set, and compare outputs side by side. The Higgsfield changelog (read 2026-10-05) dates Seedream 5.0 Flash to Sep 23 and describes it as a fast tier with multi-reference style blending for generate and edit. That is a vendor description, not a measured result, so the plan below is how you check it for your own brand board.

What the sources say

Keep the claims separate from the tests.

Five tests

Run these in order. Each changes one thing.

• Baseline: one prompt, no references. This shows what the model does by itself.
• One reference: add your single best style image. Note which attributes carry over (colour, texture, line).
• Three references: add two more from different assets. Check whether the output blends them or copies one.
• Six references: add duplicates of the same palette in different scenes. Check for drift toward the majority.
• Order swap: reverse the reference order with the same set. If the output changes a lot, order matters to your workflow, so fix it in code.

Run the same test on Sume

Flash is not in the Sume catalog. The closest catalog row is bytedance-seed/seedream-5-lite, which takes up to 10 references. The request below sends three. Replace the URLs with your own public HTTPS images. A 200 returns images in data; a 202 returns a job whose result you read from result_url.

import asyncio, os, httpx

REFS = ["https://example.com/board-1.jpg", "https://example.com/board-2.jpg", "https://example.com/board-3.jpg"]

async def main():
    body = {
        "model": "bytedance-seed/seedream-5-lite",
        "prompt": "a ceramic mug on a wooden table, in the style of the references",
        "input_references": [{"type": "image_url", "image_url": {"url": u}} for u in REFS],
    }
    headers = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
    async with httpx.AsyncClient(timeout=60) as c:
        r = await c.post("https://api.sume.com/v1/images", headers=headers, json=body)
    if r.status_code == 200:
        print([d["url"] for d in r.json()["data"]], r.json()["usage"]["cost"])
    elif r.status_code == 202:
        print(r.json()["data"]["result_url"])
    else:
        print(r.status_code, r.text)

asyncio.run(main())

Scoring

Score each output on three fixed questions: does it keep the subject from the prompt, which reference attributes show up, and would a designer accept it without edits. Log the usage.cost from each Sume call so the test also gives you a cost per keeper. Our reference-count comparison covers the 10 versus 14 limit if your board is larger.

What to do with the results

Write the winning reference count and order into your pipeline config, so the next run does not rediscover it. If the three-reference result beats the six-reference result, a higher reference cap will not help you. If six wins, the extra references are doing work, so test your largest realistic set next. Keep a note of the dates: we read the Flash page on 2026-10-05, and a model can change after launch.