Answer: a chained edit is an image-to-image job where the reference is the previous output rather than the original. Seedream 5.5's published study runs twenty of them — a coastal station from morning light to blue hour — and the interesting part is not the feature, it is the drift: each step re-renders, so colour, edges and small details move a little every time.
Twenty edits is a long chain. Drift compounds, and the difference between a chain that stays coherent and one that slowly becomes a different picture is decided by how much you change per step.
The chain, as published
The study's steps are small and additive, in this order: red tram, jade kiosk, striped awning, copper clock rim, roof tiles, platform lights, three travellers, station sign, luggage crates, ivory sail, dock ladder, bridge railing, greenhouse flowers, ferry banner, two umbrellas, blue-hour sky, wet platform, lit tram windows, bridge lights, breaking surf.
Two design lessons sit in that list:
- One object or one condition per edit. "Red tram" is a colour decision on one element. "Blue-hour sky" is a light decision affecting the whole frame. They are not mixed in the same step.
- Environment changes come after object changes. Props and signage first, then weather and time of day — because a light change resamples everything, and doing it early means every later edit inherits the new palette.
What drifts, and why
| Drift | Cause | Mitigation |
|---|---|---|
| Colour temperature creeps | each re-render re-decides white balance | name the light source in every prompt |
| Edges soften | repeated resampling | work at 1.5K or 2K, not 1K |
| Small props vanish | the model re-reads the scene each step | restate the elements that must persist |
| Composition shifts | no anchored base | keep a "do not move" clause |
The middle two are the ones people notice late: by edit ten, the awning is a different stripe count and the sign has new letters.
The prompt pattern for a long chain
Base: @Image1 (the previous edit).
Change: [one thing].
Keep: the framing, the camera height, the palette, the three travellers
on the platform, and the station sign's wording.
Output: 1.5K, 16:9.Three fields: what changes, what is frozen, what size. The frozen list grows as the chain grows — by edit fifteen it is the longest part of the prompt, and that is correct.
Chained edits versus one big edit
| Approach | Result | Cost |
|---|---|---|
| Five changes in one prompt | five compromises, one render | one edit |
| Five chained edits | five decisions, five renders | five edits |
| Chained at 1.5K | stable, cheaper | usage-based |
A single prompt asking for a tram, a kiosk, an awning, a clock and new tiles produces an image where all five are approximate. The chain produces five deliberate states — and lets you stop at any point and export.
When the chain is the wrong tool
- The composition has to change. Start a new generation; chaining cannot relayout a frame.
- Text must be exact. Isolate it with layer decomposition and set the type, rather than editing it five times.
- The deliverable is one image. Draft with three edits in a chain, then rebuild the winning state as a single render from the original plus three references.
Cost discipline
Each step is an edit, settled on usage. Twenty steps at 1.5K is still a small bill in absolute terms — a few tens of credits — but it is also twenty chances to churn. The habit that keeps it cheap: decide the entire change list before step one, so the chain is a plan rather than a search.
FAQ
Does each edit need the original as well as the previous result? Previous result only; ten references are allowed if you want to add material.
Can I branch the chain? Yes — export any step and continue from it in a second chain. That is how you compare two directions without redoing the first ten.
Does resolution cap the chain? 1K, 1.5K or 2K per step; 2K for final states.
Is the drift avoidable entirely? No — it is inherent to re-rendering. It is manageable with frozen elements and higher resolution, not removable.
The published twenty-step study is on Callirra's Seedream 5.5 page, with a comparison slider for any two states — the fastest way to see what ten edits actually do to an image.