Best Sume image model for text in images: five catalog ids compared

No vendor page proves which image model renders text best, so test on your own copy. What the vendors claim is public: Recraft's page says Ideogram excels at typography, Black Forest Labs says FLUX.2 [flex] is optimized for typography, Alibaba says Qwen-Image 2.0 Pro has stronger text rendering, and Recraft says Seedream has accurate text rendering. On Sume you can call ideogram/ideogram-v3, black-forest-labs/flux.2-flex, recraft/recraft-v4, qwen/qwen-image and the Seedream ids with the same request body and compare.

Claims are from Recraft Studio's model list, the FLUX.2 overview and Alibaba's Qwen-Image page, all read 2026-10-02. Sume's ids come from the Image API.

What does each vendor say about text?

These are vendor marketing statements, not measurements, and none of the pages gave a benchmark number.

What is a fair test?

Pick ten strings you really ship: a price badge, a two-line headline, a non-English phrase, an all-caps word and a long sentence. Use the same prompt, aspect ratio and quality across models, request one image each, and score by eye with a checklist: spelling, kerning, stray glyphs. Repeat on a second set so one lucky seed does not decide. Sume does not serve seed in v1, so you cannot lock randomness; repeat runs instead.

Set quality only where the model lists it. The Image API page says quality values are catalog-gated and that quality escalation helps for finals and dense text, so try a higher value on the model that is close.

How do you run the comparison in one script?

The script submits the same prompt to several models at once and prints what came back. Status 200 means the image is in the body, 202 means a job to poll, as the Image API page documents.

import asyncio, os, requests

URL = "https://api.sume.com/v1/images"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
MODELS = ["ideogram/ideogram-v3", "black-forest-labs/flux.2-flex", "recraft/recraft-v4", "qwen/qwen-image"]
PROMPT = "A cafe chalkboard that reads: SOUP OF THE DAY 6.50"

def one(m):
    r = requests.post(URL, headers=H, json={"model": m, "prompt": PROMPT}, timeout=60)
    return m, r.status_code, r.json()

async def main():
    for m, code, body in await asyncio.gather(*(asyncio.to_thread(one, x) for x in MODELS)):
        if code == 200:
            print(m, body["data"][0]["url"], body["usage"]["cost"])
        else:
            print(m, code, body)

asyncio.run(main())

What if you do not want to choose?

Send model: "sume/auto" and let the Image Router choose; Sume does not disclose which family ran. That trades control for convenience, which is fine for drafts and wrong for a typography A/B test. For picking in practice see how to pick a model for product photos, and for FLUX flex's text behaviour see FLUX.2 flex text rendering.

How to record the result

The honest takeaway is that text rendering is a per-prompt property. A model that nails an English headline can still fail on a Korean phrase or a price with decimals, so keep your own test set in version control.

• Save the model id, prompt, aspect ratio and quality next to every output.
• Score spelling first, layout second, style third.
• Rerun the winner three times to see how stable it is.
• Re-test when the catalog changes; a newer id may beat last month's winner.