Answer: Seedream 5.5's image-to-image mode accepts up to ten reference images from your library, at up to 30 MB each, in formats from PNG and JPEG through WebP, BMP, TIFF and HEIC. The base image sets the composition; the other references set what to keep or introduce. Past four or five, references start averaging unless each one is given a distinct job.
Ten is a ceiling, not a plan. The edits that work use two to four references and a prompt that says which is which.
What the inputs accept
| Limit | |
|---|---|
| References per edit | 10 |
| File size each | 30 MB |
| Formats | png, jpg, jpeg, webp, bmp, tif, tiff, gif, heic, heif |
| Source | your own library |
| Prompt | 5,000 characters |
| Output | PNG or JPEG, 1K / 1.5K / 2K |
Because the references come from the library rather than a URL, the practical workflow is: upload once, reuse across edits. That also makes chained edits cheap in attention — the same base image can be edited twenty times without re-uploading.
Assigning roles
The failure mode with many references is that the model treats them as a mood board. The fix is to make each reference own one decision, and say so in the prompt:
Base: @Image1 — keep the framing, the horizon and the light direction.
Wardrobe: @Image2 — the coat, same fabric, same colour.
Prop: @Image3 — replace the crate with this instrument case.
Everything else: unchanged.Three references, three jobs, one sentence each. That prompt is more reliable than ten references and the word "similar".
What edits can and cannot do
Can: change light and time of day, swap wardrobe or props, restage an object on a different surface, add or remove a figure, restyle the render (painterly to photographic), and correct a composition without re-prompting the scene.
Cannot: recover detail that was never in the base image, guarantee text accuracy (see the layer mode's text recognition), or preserve exact pixel geometry — an edit re-renders, so expect small shifts in edges.
If the deliverable needs the same pixels with one element changed, that is a compositing job, not a generation job.
The prompt habits that matter
- Name the base. "Keep the composition of @Image1" is the anchor; without it the model may regenerate the scene.
- Give each reference a noun. Wardrobe, prop, surface, light, palette — not "style reference".
- State what stays. A single line, "everything else unchanged", prevents collateral drift.
- One change per edit. Chaining five changes in one prompt produces five compromises; chained edits produce five decisions.
Cost, per edit
Edits settle on usage with a credit ceiling rather than the flat per-image price. A single-reference touch-up sits near the 2–4 credit range; a ten-reference restyle with a large output costs more, because the charge includes the references and the delivered outputs.
| Edit | Approximate credits |
|---|---|
| 1K, one reference, one change | 2–4 |
| 2K, three references, restyle | 4–8 |
| 2K, layer split (see below) | usage-based |
Check the quote before submitting; the ceiling exists so a runaway edit cannot exceed it.
When to use a different mode
- You need editable parts. Layer decomposition, not editing.
- You need a new image from a description. Text to image.
- You need exact type. Render the layout, then check the letters; the model is good with text but not typesetting.
FAQ
Can I reference an image I generated in another session? Yes — anything in your library.
Do references cost extra? They are included in the usage-based settlement, which is why the ceiling exists.
Is there a way to lock the seed? No seed control is exposed; consistency comes from chaining edits rather than from locking a number.
Does the base image have to be 1:1? No, but output ratio follows the job setting — the eleven-ratio set includes 21:9 and 1:2.
The mode and its limits are documented on Callirra's Seedream 5.5 page, where the twenty-edit study shows the same image revised twenty times from a single base.