Back to Blog
Guides
7 min read

How to Write Seedream 5.5 Edit Prompts

A practical method for image-to-image prompts on Seedream 5.5: change first, then what stays, one step per round, plus a five-round worked example.

ByteDance's newest Seedream model is built for editing in rounds: change one thing, look, change the next. Its release notes say regions you don't touch are preserved at the pixel level, so a picture can survive many passes without going soft or drifting in color. That only pays off if each prompt tells the model clearly what to change and what to leave alone. This guide shows how we write those prompts on this site.

1. Lead with the change, then list what stays

Put the action first, in one sentence, and the "keep" list right after it.

Replace the paper cup with a white ceramic mug. Keep the hand, the laptop and the window light exactly the same.

Two short sentences beat one long one. The model reads the first as the job and the second as the constraint. Vague words like "improve" or "make it better" give it nothing to aim at; say what better looks like ("brighter, warmer, less cluttered").

2. One or two changes per round

It is tempting to write the whole brief at once: new background, new jacket, different light, a coffee cup on the table. Every extra change gives the model more room to reinterpret the picture. Make one or two changes, check the result, then upload it and continue. On Seedream 5.5 the untouched parts should hold steady from round to round, which is exactly what makes this approach work.

Each round is a new image and is billed as one, so draft the big moves on Seedream 5.0 Flash (2 credits) if you are still exploring.

3. Point at things the way the picture shows them

The editor here works from words alone, so the location has to be in the prompt. Name the object by what makes it unique in the frame:

  • "the blue mug on the left", not "the mug"
  • "the poster above the sofa", not "the poster"
  • "the woman in the yellow coat, second from the right"

If two objects look alike, add position and color. If the model changes the wrong one, the reference was ambiguous; tighten it rather than repeating the same prompt.

4. Refer to uploads by number

Seedream 5.5 takes up to 4 reference pictures here. They are numbered in upload order, so your prompt can say exactly what each one contributes:

Put the lamp from Image 2 on the desk in Image 1, on the right side next to the books. Match the warm evening light of Image 1. Keep everything else in Image 1 unchanged.

Upload the picture you are editing first. That way "Image 1" is always the base, and Auto keeps its shape.

5. Describe light and material, not just objects

Realism is one of the model's strengths, and it responds well to physical detail. "A linen sofa" looks better than "a sofa"; "late-afternoon sun from the left, long soft shadows" looks better than "nice lighting". When you add something new, tell the model how it should sit in the scene: "casting a soft shadow on the table", "reflected in the window".

6. Put words in quotes

For text on signs, labels or packaging, quote the exact words and say where they go and how large:

Change the shop sign to read "ROSIE'S BAKERY" in the same font and color. Keep the window display and the street unchanged.

ByteDance reports that 5.5 renders dense small text more cleanly than earlier versions. For whole paragraphs or a second language, though, Seedream 5.0 Pro is still the safer choice.

A worked example: five rounds on one photo

Starting picture: a plain home office, white walls, a bare desk, a window on the left. Send each prompt below with the previous result as Image 1.

  1. "Paint the wall behind the desk a muted sage green. Keep the desk, the window and the floor exactly the same."
  2. "Add a tall potted fiddle-leaf fig in the right corner, on the floor, slightly behind the desk. Keep everything else unchanged."
  3. "Put a walnut desk lamp on the left side of the desk, switched on, with a warm pool of light on the desktop. Keep the plant and the wall color."
  4. "Hang a large framed black-and-white landscape photograph on the green wall, centered above the desk. Keep everything else the same."
  5. "Make it early evening: the window shows a dusky blue sky, the desk lamp is the main light. Keep all furniture, the plant and the frame exactly as they are."

Notice the pattern: one addition per round, a clear position for every new object, and a reminder of what was added earlier so it stays put.

When an edit goes wrong

  • The model changed something you didn't mention. Add it to the "keep" sentence explicitly, then run the round again from the previous good image, not from the bad one.
  • The new object looks pasted on. Describe its light and shadow, and its contact with the surface ("resting on", "leaning against").
  • A face drifted. Add "keep the face, hair and expression exactly the same", and avoid changing pose and expression in the same round.
  • Text came out wrong. Shorten it, quote it, and say where it goes. For long copy, switch to Seedream 5.0 Pro.

Templates you can copy

Replace [object] with [new object]. Keep [the subject], [the background] and [the light] exactly the same.

Add [object] [exact position], [size relative to something in the frame], [how light falls on it]. Keep everything else unchanged.

Put the [item] from Image 2 on the [subject] in Image 1. Match the lighting of Image 1. Keep [face / pose / background] unchanged.

Change the text on the [sign / label] to "[EXACT WORDS]" in the same font and color. Keep everything else the same.

Try them on the Seedream 5.5 page, or read how the three Seedream models compare before you choose one.

Source: ByteDance Seed, Introducing Seedream 5.5 (November 9, 2026).