Seedream 4.5: change one locked still (4K, 4cr)

Seedream 4.5 is an edit. Point at the object. A new text-to-image throws away the still you approved.

Seedream 4.5 is ByteDance's all-round image model: generation, editing, and image-to-image in one endpoint, with the strong text rendering and prompt comprehension the Seedream line is known for. Four credits. 4K. Categories: text-to-image, edit-image, image-to-image. That third and second category are why this page exists. If the still is already signed, do not roll a new generate and hope the lettering comes back.

One endpoint is not a license to regenerate

All-round is a convenience, not a workflow. The same four credits can create a still or change one. Those are different jobs. Creating again throws away the approval. Editing keeps the structure and moves one object, one color, one word.

The catalog does not require an image (requires_image is false) because the endpoint also generates. That is a trap if you live in "just prompt it again." Prompting it again is how the approved bottle becomes a cousin of the bottle. Point at the object instead.

ByteDance's video family is Seedance; this is Seedream. The split is on the ByteDance roster. Do not send a locked still into a video model to "fix the label." That is a different content type.

Point at the object

An instruction edit names the thing and the change: the parked car, the sky, the word on the box. Everything you did not name stays. That is the image-to-image contract, and it is why Seedream 4.5's text rendering matters — posters, packs, and UI mockups live or die on the letters you already approved.

Use the AI photo editor when the file exists. Describe one change. If you need three changes, run three four-credit edits rather than one prompt that asks the model to reinvent the room. Inpainting language ("this region, this replacement") beats "make it better."

If you do not have a file yet, then yes, the text-to-image side of the endpoint is fair. Generate, approve, then edit. Do not skip the middle.

4K text you already signed

Quote the row:

• 4 credits
• 4K max output
• text-to-image + edit-image + image-to-image
• no durations, no audio (it is an image model)
• up to 10 reference slots on max_reference

Aspect ratios include 1:1, 2:3, 3:2, 4:3, 3:4, 16:9, 9:16, 21:9. Keep the ratio of the approved still. A 4K edit that silently reframes 9:16 into 1:1 is not an edit. It is a new picture.

Strong text rendering is the reason to stay on this line when the still has words. A generic generate will happily redraw the type. Seedream 4.5 is the row that is supposed to understand the prompt and the letters. Use that. Do not gamble it on a fresh roll.

Four credits to change, not to throw away

Four credits feels like "one more try." That feeling is how approved work dies. If the still is locked, the cheap move is an instruction, not a new sample. The ranked editing list is best image editing model if you are choosing among edit rows. This page is the Seedream 4.5 rule: same file, one object, 4K out.

When the edit is done, then you may buy motion. Animate the locked still. Do not animate a regenerate that almost matches.

FAQ

Can I use Seedream 4.5 as a plain text-to-image model?

Yes. The categories include text-to-image, and requires_image is false. The failure mode is using that generate path after a still is approved. Once it is signed, switch to the edit.

Why not regenerate at 4K until the label is perfect?

Because each generate is a new picture. Text rendering is a reason to edit the one that already has the right layout, not a reason to roll dice until a cousin appears.

Does this row make video?

No. No durations, no audio, image content type. Motion is a later catalog line. Keep the 4K still; animate it elsewhere.

What do I actually type?

Name the object and the change. "Replace the headline with SUMMER RESTOCK, keep the rest." Not "a better poster of the same campaign." The second prompt is a generate wearing an edit costume.