200 variants of one photo for a test set: augmentation cost on Sume

Two hundred edited variants of one product photo (new backgrounds, lighting, angles) take 50 calls with n: 4 and cost $5.00 on Qwen Image, $6.50 on Seedream 4.0 or $7.50 on Flux 2 Pro on Sume. That is the full bill for a small evaluation or demo set.

All three models take the source photo as a reference and allow n up to 4 (read 2026-10-03). Label the output as synthetic wherever it is stored or shared; generated variants should never be mixed into a real-photo evaluation set without a tag.

Cost of 200 variants

The table below lists the figures this post uses, as read on 2026-10-03.

Making the variants useful

Vary one factor per batch: background in one run, lighting in the next, camera angle in a third. A prompt template with a single changing slot makes the set auditable, and you can store the slot value next to each result.

The metadata field is stored on the job and not sent to the provider, so it is the right place for the batch name, the slot value and the source photo id. When a variant turns out to be wrong, the metadata tells you which prompt produced it.

• Check each variant against the original for object identity: logos, labels and shapes should not change.
• Drop near-duplicates before you count the set; four samples of one prompt are often similar.
• Keep generated files in a folder separate from real photos.

One call of the loop

Use aspect_ratio: "auto" only on models that list it. Seedream 4.0 does; Qwen Image and Flux 2 Pro do not, so name the ratio of your source photo instead.

{
  "model": "bytedance-seed/seedream-4",
  "prompt": "Same product, new background: {slot}. Keep the label unchanged",
  "aspect_ratio": "auto",
  "n": 4,
  "metadata": {"batch": "aug-01", "slot": "kitchen counter"},
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/source.jpg"}}
  ]
}