Multi-turn ad edits with Ideogram 4.5: four rounds for 30 cents

You can chain Ideogram 4.5 edits on Sume by passing each result URL as the input reference of the next call. Four medium-quality rounds cost $0.30, and four high-quality rounds cost $1.10, at Sume's list-times-1.25 rates.

Ideogram's launch post says 4.5 eliminates artifact buildup, which is what makes multi-turn editing possible. That is the vendor's claim. The cost math below is Sume's, and a short test on your own banner is the way to check the claim.

What the vendor says, and what Sume documents

In the Ideogram X post (read 2026-10-05), the company describes 4.5 as the most precise edit model, live in Ideogram, the API and launch partners, with open weights coming. It says nothing in that post about price or limits. A Morphic page (read 2026-10-05; a partner, not the vendor) lists a 25 MB file limit, an aspect range of 1:6 to 6:1, and mask support. Sume's docs do not list those, so I do not rely on them. What Sume documents is the shape of the call: with input_references the model edits the first image and uses up to four more as references.

The arithmetic

Sume's Image API docs give fal list prices of $0.03, $0.06 and $0.22 per image by quality, which Sume bills at list times 1.25. That is $0.0375, $0.075 and $0.275.

A four-round holiday edit

A good habit is to change one thing per round and say what must not change. A holiday ad might go like this.

• Round 1: swap the product color to the seasonal edition.
• Round 2: change the headline to the sale line.
• Round 3: add a small 'Ends Monday' badge in the bottom-right corner.
• Round 4: crop the safe area for a 9:16 placement by setting aspect_ratio.

Draft low, finish high

Draft the early rounds at low, since layout and wording are easy to judge at that quality, and run the last round at high. That plan costs three low rounds and one high round: $0.1125 + $0.275 = $0.3875, under $0.40, and the final pass gets the best quality. Use resolution: 2K only for the last pass if the placement needs it.

Keep every intermediate file. Sume returns Sume-hosted, signed URLs in data[].url. Reference URLs must be public HTTPS, so if a signed URL is refused as an input, download the image and host it yourself before the next round. If round three goes wrong, restart from round two's output and you lose one edit, not four.

A chained edit can still drift. Compare each round to the source by eye. If text you did not touch has changed, restart from the last good image and tighten the instruction. The prompt should name what to preserve as clearly as what to change.

After the last round

Then animate the winner, or hand it to the ad platform as a static. A still costs a fraction of a clip, so a still ad variant is the cheapest way to test an offer before you commit to video.

How to test the claim

The reason to care about multi-turn editing is that real ad work is a chain of small changes, not one perfect prompt. A merchandiser asks for a different color, then a shorter headline, then a badge, then a crop. With an edit model that degrades the image each time, you stop after two rounds and start over. With one that holds the untouched pixels steady, you keep going.

That is the claim to test. Take one banner, run the four rounds above, and set the first and the fourth image side by side. Look at the areas you never touched: the logo, the product edges, the small print. If they match, the chain is safe for your work. If they drift, cut the chain to two rounds, or generate the final from a clean source with a combined instruction.

Also test the instruction style. Short imperative lines with an explicit 'leave everything else unchanged' tend to be easier to review than long descriptions, because you know what the edit was supposed to do. Write the instruction as if a colleague will check the result against it.

Budgeting the batch

The cost of a test is small enough to run before every campaign. Ten banners through four medium rounds is $3.00. Ten banners through three low rounds and one high round is $3.875. Either fits inside a hook-test budget, and either gives you a stack of approved statics to pair with your video variants.