Multi-turn edits or one combined prompt: Ideogram 4.5 restart rule

Chain edits when each change depends on the last one, and restart from the original with one combined prompt when the changes are independent or when the image starts to look worn. Ideogram's launch post (read 2026-10-05) says 4.5 "eliminates artifact buildup, making multi-turn editing possible", so chaining is a supported pattern. It is not the only one, and the claim is the vendor's, not a measurement of your images.

A decision table

What drift looks like

Even if a vendor says buildup is solved, look for the signs yourself on long chains. The usual ones are soft edges on text that was crisp, slight color shifts in large flat areas, faces or hands that change shape from pass to pass, and small objects that move. None of these is proven to occur or not occur for Ideogram 4.5; they are what to look for in any chain. The first pass that shows one is the pass to restart from.

A written log helps. For each pass, store the prompt, the job id, the result URL and a one-line verdict ("text ok", "edge soft"). When a late pass looks wrong, you can return to the last pass with a good verdict and branch from there.

Cost of each pattern

A chain of N passes costs N billed images. One combined prompt costs one billed image per attempt, but the attempt may need more tries because it asks for more at once. A restart costs the new attempts again. There is no rule that wins on cost in general, which is why a pilot on your own images is worth running. Use quality: low for pilots, then read the per-image line from GET /v1/images/models/ideogram/ideogram-v4.5/endpoints for the real budget.

How Sume handles each turn

Sume's Image API is stateless: each call is a new request that carries its own input_references. A "turn" is therefore a new POST /v1/images whose first reference is the previous result URL. The Image API docs say with references the model edits the first image and takes up to 4 more as references, 5 in total.

Each pass is its own job and its own billed image. A chain of four passes is four generations, whatever the vendor says about buildup. Use a webhook or poll status_url between passes so pass two never starts with a half-finished pass one.

A test you can run today

Take one real image. Run a chain of three different edits. Then run one combined prompt that asks for all three changes. Compare both against the original at 100 percent zoom, looking at text edges, faces and flat color areas. If the chain matches the combined result, you can pick on convenience and cost. If it does not, note which pass introduced the difference and restart from the original there.

Keep the prompts short: one change each. Say what must stay identical every time. Record the job id, prompt and result URL for each pass, so you can go back to the last good one.