ByteDance Seed released Seedream 5.5 on November 9. Its pitch isn't "prettier pictures". It's "keep editing the same picture without it falling apart". The headline claim: up to 100 rounds of edits on one image, with the parts you didn't touch preserved at the pixel level and no color drift (ByteDance Seed).

The launch landed. ByteDance Seed's announcement on X passed 370,000 views, and CapCut's post about the model drew close to 400,000 views and more than 4,700 likes within two days (ByteDance Seed on X, CapCut on X). On November 10, US Google searches for "seedream" were running at more than twice their level of a week earlier (Google Trends).

This post covers what's new, which claims are still unverified, and three prompts that test the same skills. We ran every prompt on Seedream 5 Pro, the Seedream model you can use on EasyImagen today, so you can see where the 5.x family already stands.

What's new in Seedream 5.5

From ByteDance Seed's announcement and The Decoder's report:

  • Stable multi-round editing. You can keep working on the same image for up to 100 rounds. Unedited areas stay pixel-identical, and detail loss across successive edits is reduced. The demo edits one portrait round after round, swapping clothes, rings and glasses.
  • Native transparent output. The model can return RGBA images with full or partial transparency, so stickers, product cutouts and text assets don't need a separate background-removal step.
  • More natural realism. Lighting, materials and skin look less glossy and less overexposed.
  • Dense small text. ByteDance reports the rate of severe text corruption fell to 0.4% and the minor rate to 1.7%.
  • Where it runs. It's rolling out in ByteDance's Jimeng, Doubao and Doubao Work apps and on Volcengine Ark, and ByteDance links to Dola and a BytePlus playground for trying it outside China. A Pro version of the 5.5 series is planned. Krea and Magnific announced support on November 10 (Krea on X, Magnific on X).

What's still unproven

Every number above comes from ByteDance. Its quality scores were judged by expert evaluators on its own MagicArena platform. The rival models it names, GPT Image 2.5 Sunburst and Nano Banana 2.1, aren't the newest ones. OpenAI's GPT Image 2.6 showed up on Arena five days before the launch, as XenoSpectrum points out. The 100-round demo is also a single portrait, with no data on how quality changes as edits pile up. Treat "100 rounds" as a design goal until independent tests confirm it.

How to prompt for multi-round edits

Even on a model built for it, edits drift faster when the prompt is vague. Three habits keep an image stable for longer:

  1. Change one thing per round. "Swap the flowers and change the sign" invites the model to redraw both areas and everything between them.
  2. Name what must not change. List the important objects, any visible text word for word, and the framing and lighting. A plain "keep everything else the same" helps, but naming things helps more.
  3. Always edit the latest version. Feed each round's output back in as the reference image. That's the workflow Seedream 5.5 is designed for.

We tested this on a bakery window, a typical small-business photo that needs seasonal updates. First, the starting image:

A photograph of a small neighborhood bakery's front window on an early winter morning, seen from the sidewalk. Inside the window: a wooden shelf with round sourdough loaves and a few baguettes in a wicker basket, a clear glass vase of yellow tulips on the left, and a black chalkboard sign on a small easel in the center with hand-lettered white chalk text that reads exactly "FRESH SOURDOUGH" on the first line and "Open 7am – 3pm" on the second line. Soft warm light from inside the shop, cool blue daylight outside, faint reflections on the glass, a brick wall edge on the right. Realistic, natural colors, 35mm photo. No other text, no logos.

A bakery window with sourdough loaves and baguettes on a wooden shelf, yellow tulips in a glass vase and a chalkboard reading FRESH SOURDOUGH, Open 7am – 3pm

Round 1: swap one prop for the season. We passed the image above in as the reference.

Edit only the vase on the left: replace the yellow tulips with a few fresh pine branches and red winterberry sprigs in the same clear glass vase, and add one strand of tiny warm fairy lights draped along the front edge of the wooden shelf. Keep everything else exactly the same: the loaves, the baguette basket, the chalkboard sign and its text "FRESH SOURDOUGH" / "Open 7am – 3pm", the easel, the window frame, the brick wall, the reflections, the framing and the lighting.

The same bakery window, now with pine branches and red winterberries in the vase and a thin strand of fairy lights along the shelf

The loaves, sign, easel and reflections came back unchanged, and the chalk lettering is identical. One miss: the fairy lights are there but barely glowing. If you want them lit, say so ("glowing warm fairy lights").

Round 2: change the text, nothing else. This time the reference was the round 1 output.

Edit only the chalk text on the chalkboard sign. Change the first line to "HOLIDAY STOLLEN" and the second line to "Order by Dec 20", in the same hand-lettered white chalk style, size and position, with the same faint chalk smudges. Keep everything else exactly the same: the pine and winterberry vase, the fairy lights on the shelf, the loaves, the baguette basket, the easel, the window frame, the brick wall, the reflections, the framing and the lighting.

The same bakery window with the chalkboard now reading HOLIDAY STOLLEN, Order by Dec 20, and the pine vase and fairy lights unchanged

The new lines match the old chalk style, and the round 1 changes carried through. Two rounds is a long way from 100, but this is the habit that makes long edit chains work on any model. For more edit prompts, see our MAI Image 2.7 editing guide.

Testing small text

ByteDance's text demos are a newspaper page and an illustrated infographic. To test text yourself, spell out every word you want in quotes and keep each caption short:

A clean editorial infographic poster titled "POUR-OVER COFFEE IN 6 STEPS" in bold black sans-serif at the top, with the subtitle "20 g coffee · 320 g water · 3 minutes" in smaller type underneath. Below, a 2-column by 3-row grid of six panels on warm off-white paper. Each panel has a simple flat illustration in muted terracotta, sage and charcoal, a large step number and a short caption in small type, exactly:
1 "Heat water to 94 °C"
2 "Grind 20 g, medium-fine"
3 "Rinse the paper filter"
4 "Bloom with 40 g for 30 s"
5 "Pour slowly to 320 g"
6 "Drain by 3:00, then serve"
A thin rule separates the panels. At the bottom a small footer line reads "Ratio 1:16". Crisp vector-like print design, generous margins, every word spelled exactly as written, no other text, no logos.

An infographic poster titled POUR-OVER COFFEE IN 6 STEPS with six illustrated panels showing a kettle, grinder, filter, bloom, pour and finished carafe, each with a short caption

Seedream 5 Pro got the title, subtitle, all six captions and the footer right, including the degree sign and "1:16". Where models still slip is long paragraphs at very small sizes. That's the case ByteDance says 5.5 improves, and the one to check first when you get access.

Testing realism

The cover image is our realism test. It uses backlight, dust in the air, wet clay and aged skin, which are all places where models tend to over-smooth or over-brighten:

A documentary-style photograph of a ceramicist in her sixties trimming a bowl on a potter's wheel in a small studio, late afternoon sun coming through a dusty window behind her. Backlight rims her grey hair and the edge of the spinning clay, fine clay dust floating in the light, wet clay slip on her fingers and forearms, natural skin texture with fine wrinkles and no retouching, a canvas apron with dried clay marks, shelves of unglazed bowls softly out of focus. Natural exposure that keeps detail in the bright window and the shadows, no glossy skin, no oversaturation. Shot on a 50mm lens at f/2, muted warm color. No text, no logos.

Phrases like "no glossy skin", "no retouching" and "keeps detail in the bright window" target exactly the problems ByteDance says 5.5 fixes. They help on today's models too.

Try it now

Seedream 5.5 isn't on EasyImagen yet. Seedream 5 Pro is, and so are GPT Image 2 and Nano Banana, if you want to run these prompts side by side. Open the generator, paste a prompt, and for edits upload your previous result as the reference image. For a holiday project, try our AI Christmas card generator, or try this week's miniature Christmas village prompt.