Repeated image edits without drift: a native-resolution test plan

To test whether an image model drifts under repeated edits, run the same small change five times in a row, always feeding the previous output back in, and compare the first and last results at 100 percent zoom. Recraft's October 1 post says Ideogram 4.5 was built for exactly this case: stable repeated edits, with edits done at native resolution without downsampling.

What Recraft says

The Recraft post states that Ideogram 4.5 targets stable repeated edits and performs native-resolution edits without downsampling, and that it is coming to Recraft Studio. It does not publish a drift measurement, so the claim is a design goal to check, not a measured guarantee.

The five-pass protocol

Pick one product photo and one change per pass, small and specific, such as swapping a label colour or changing one word of text. Keep all other settings fixed.

Running it on Sume

Sume's Image API lists Ideogram 4.5 as ideogram/ideogram-v4.5. Without input_references it generates from text. With them it edits the first image and uses up to four more as references, five in total. An edit without aspect_ratio keeps the source image's shape, and resolution takes 1K or 2K. Quality is low, medium or high, defaulting to medium.

Reference URLs must be public HTTPS. Use the Sume URL returned from each pass as the input to the next.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ideogram/ideogram-v4.5",
    "prompt": "Change only the label colour to deep green; keep everything else identical",
    "quality": "medium",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/product.png" } }
    ]
  }'

Cost of the test

The docs give the Fal list price for Ideogram 4.5 as $0.03, $0.06 or $0.22 per image by quality, whatever the size. Five passes at medium come to $0.30 at list, at low $0.15 and at high $1.10. The price you pay is the endpoint pricing line in the catalog, so read it before a large run. Failed generations are not billed.

Compare pass 5 with pass 1 at 100 percent zoom. If text and edges hold, the model passes your test for that kind of image.