Seedream Edit: change one locked still (4K, 3cr)

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

Seedream Edit is ByteDance's instruction-driven change to an existing image: replace objects, adjust backgrounds, fix details. The catalog calls it the budget-tier edit option in the Seedream family. Categories: edit-image and image-to-image. Requires image: yes. 4K. 3 credits. The AI photo editor is the surface. A fresh generate is the other surface, and it does not owe you the face you already signed off.

The input is the contract

No still, no job. The row will not invent the hero from a paragraph and then "keep it consistent." You upload the approved frame. You write the change: swap the cup, clean the reflection, move the sky. Everything you did not name should survive.

That is image-to-image, not a second roll of the same prompt. Re-rolling text-to-image is how a locked label becomes a cousin of the label. Cousins fail legal. Cousins fail the grid.

Max reference on the snapshot is 10. Use references for the thing that must stay, not as a collage of vibes. One locked still plus a pointed instruction is the happy path.

Budget-tier means 3 credits, not "sloppy"

Description, quoted: ByteDance's Seedream editor — instruction-driven changes to an existing image: replace objects, adjust backgrounds, fix details. The budget-tier edit option in the Seedream family.

3 credits. The row is marked discounted in the snapshot; do not treat discount as a license to spray ten variants of the same sky. Each edit is a chance to drift. Drift is the failure mode of cheap iteration.

Output: 4K (max_output_resolution). Qualities: SD, HD, 4K. Aspects: 1:1, 2:3, 3:2, 16:9, 9:16. Match the plate you uploaded. Do not "recompose for 9:16" inside an edit unless the crop is the instruction. Recompose is how you lose the product at the edge.

Audio is null. Durations are empty. This is a still. If you needed motion, you are on the wrong ByteDance line — Seedance is video, and it lives on the ByteDance roster as a different family.

What you throw away if you start over

A new text-to-image does not "fix the hand." It samples a new person. The approved still is a legal artifact: wardrobe, SKU, talent release, the crop the ads account already ran. Seedream Edit is the tool that respects that artifact.

Point at the object.

• Wrong bottle cap → name the cap.
• Busy background → name the background.
• Soft type on a sign → say so, knowing type is still the risky part of any edit.

Then stop. The ByteDance roster is the family map; the editor surface is the AI photo editor. This page is the 3-credit 4K Seedream row, not the whole family.

If the still was never good, do not edit it. Generate a new hero on a text-to-image row and lock that. Editing a bad plate is how you spend 3 credits to keep a bad plate.

FAQ

Can I use Seedream Edit without uploading an image?

No. Requires image is true. Categories are edit-image and image-to-image. A prompt with no plate is a different model.

Why not regenerate the still from text?

Because regeneration is not an edit. You lose the approved pixels. Seedream Edit is 4K, 3 credits, instruction-driven change to the file you already have.

Is this the same as Seedream text-to-image?

No. This is the budget-tier editor in the family. Replace, adjust, fix. If you need a new composition, leave this row.

Does the edit include audio or clip length?

No. Audio is empty. Durations are empty. You get a still at up to 4K. Animate after the edit is approved, on a video row.