Compare 4 image models in one labelled 2x2 sheet with cost per tile

Call POST /v1/images once per model with the same prompt, fit each result into a 512 pixel tile with ImageOps.fit, and draw the model id and usage.cost on a black strip at the bottom of the tile. Four calls with the models below cost about $0.24 in total, so you can run the sheet for every new prompt without worrying about budget.

Why a sheet beats four browser tabs

When you choose a model for a product shot, the question is rarely which one is best. It is which one is good enough at the lowest price for this prompt. A sheet puts the answer on one screen, with the price printed on the image so the screenshot you paste into a ticket carries its own evidence.

The four models here are all text-to-image capable. Recraft V4 is text-only in the Sume catalog and returns WebP, which Pillow opens without a flag, so it can sit in the same loop.

The script

Each model takes its own default size. ImageOps.fit crops to the tile ratio instead of squashing, so a wide result is cropped rather than distorted. The label uses Pillow's built-in bitmap font, which needs no font file.

import os, io, requests
from PIL import Image
H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
def gen(**body):
    r = requests.post("https://api.sume.com/v1/images", json=body, headers=H, timeout=60)
    r.raise_for_status()
    if r.status_code == 202:
        raise SystemExit("queued, read /v1/jobs/{id}/result: " + r.text)
    return r.json()
def fetch(u):
    return Image.open(io.BytesIO(requests.get(u, timeout=60).content))
from PIL import ImageDraw, ImageOps
MODELS = ["bytedance-seed/seedream-4.5", "black-forest-labs/flux.2-pro",
          "google/nano-banana-2", "recraft/recraft-v4"]
PROMPT = "Matte black espresso machine on a marble counter, morning light"
sheet = Image.new("RGB", (1024, 1024), "white")
total = 0.0
for i, m in enumerate(MODELS):
    out = gen(model=m, prompt=PROMPT)
    tile = ImageOps.fit(fetch(out["data"][0]["url"]).convert("RGB"), (512, 512))
    d = ImageDraw.Draw(tile)
    d.rectangle((0, 472, 512, 512), fill="black")
    d.text((8, 486), m + "  $" + str(out["usage"]["cost"]), fill="white")
    sheet.paste(tile, ((i % 2) * 512, (i // 2) * 512))
    total += out["usage"]["cost"]
sheet.save("sheet.png")
print("sheet cost", round(total, 4))

What each tile should cost

Planning numbers from the repo catalog at default quality and 1K output. The printed usage.cost is what you are billed.

Make the comparison fair

• Use the same prompt, and leave aspect_ratio off so every model uses its default, or set one ratio all four list.
• Run each model more than once if the decision matters. One sample hides variance, and n lets you get several takes per call where the model allows it.
• Do not compare a text-only model against an edit model on an edit task. Reference-image prompts need a model that accepts input_references.
• Record the date. Catalog prices and model behavior change, and a sheet from last quarter is not evidence this quarter.

Next steps

Once the sheet shows a clear winner, encode the choice in code with the model filter, and lock it in with a prompt regression test so a later model swap shows up as a diff. If you only need variants from one model, the contact sheet from n variants is the cheaper pattern.