Seedream 5.0 Lite edit on Sume: no auto ratio, so set the shape

When you edit a photo with Seedream 5.0 Lite on Sume, you cannot ask for aspect_ratio: "auto", because the row does not list it, so choose the ratio of your source yourself. Of the three Seedream rows in the catalog read on 2026-10-10, only Seedream 4.0 lists auto.

That is the opposite of what many users expect from a newer model. The Image API docs say a parameter outside a row's descriptor returns 400 unsupported_parameter, and the ratio enum is part of the descriptor, so auto on 5.0 Lite or 4.5 fails fast instead of guessing.

What each Seedream row lists

All three rows take up to 10 references and are billed per output image. Seedream 5.0 Lite is $0.04375, Seedream 4.5 is $0.05 and Seedream 4.0 is $0.0325. The 4.0 row lists 10 ratios including auto; 5.0 Lite and 4.5 list 9 ratios without it. None of the three lists resolution, quality, mask_url or background.

How to pick the ratio for an edit

Read the pixel size of the source and reduce it to the nearest ratio the row lists. A 4000 by 3000 photo is 4:3. A 1080 by 1350 social crop is 4:5. A 3024 by 4032 phone photo is 3:4. If the source ratio is not in the list, the edit will either crop or letterbox depending on the model, so choose the listed ratio closest to it and plan to crop afterwards.

Do this in code, not by eye. Fetch the row's aspect_ratio.values, compute width / height for each, and take the one with the smallest difference from your source. The same function works for every row that lacks auto, and it is a better habit than hard-coding a ratio per caller.

def nearest(ratios, w, h):
    target = w / h
    best = None
    for r in ratios:
        if r == "auto":
            continue
        a, b = r.split(":")
        d = abs(int(a) / int(b) - target)
        if best is None or d < best[0]:
            best = (d, r)
    return best[1]

print(nearest(["1:1", "4:3", "3:4", "16:9", "9:16"], 3024, 4032))

When 4.0 is the better edit row

If your edit workflow has varied source shapes and you do not want to handle the mapping, Seedream 4.0 is the one row in the family where auto is allowed, and it is also the cheapest at $0.0325. If you want the newer model, keep the mapping code. The choice is between a cheaper row with auto and a newer row with a small amount of glue code.

The catalog says nothing about edit quality, so judge that on your own images. Run five edits of your usual source type on 4.0 and 5.0 Lite and compare. The price difference is $0.01125 per image, so a 1,000-image run costs $11.25 more on 5.0 Lite.

• Send an explicit ratio on 5.0 Lite and 4.5.
• Send auto only on 4.0, GPT image rows and Nano Banana rows.
• Log the row and the ratio with every edit so mismatches are easy to trace.

Related checks

Because the error is a 400 and not a retry, a test that sends auto to every row you use will catch the problem in CI. Add it next to your other catalog checks.

What happens if you send auto anyway

The request is validated against the row's descriptor before any provider is called. Because auto is not in the 5.0 Lite ratio list, the call fails with a 400 and no image is generated, so you are not billed for a wasted render. That is the practical benefit of the strict descriptor: a wrong ratio costs you one failed request, not an unwanted image.

If your client library turns every 400 into an exception, catch it near the call and log the row id with the ratio you sent. A missing auto is a configuration error and retrying will not fix it.

A migration note

Teams that moved from Seedream 4.0 to 5.0 Lite for the lower price per image than 4.5 often carry the old request body forward. If that body had aspect_ratio: "auto", it needs a change. Search your code for the literal string, replace it with the computed ratio, and add a unit test that feeds a 3:4 and a 16:9 source through the helper.

Keep the helper next to the model id constant, so changing a row also reminds you to check its descriptor.