AI news4 min read
Seedream 5.5: AI image editing may never be the same
ByteDance’s new image model keeps a picture stable across 100 rounds of edits, outputs transparent PNG-style layers, and sets small text far more cleanly.
In this story6 parts
ByteDance’s Seed team has released Seedream 5.5, the latest version of its image generation and editing model. The pitch is in the title of ByteDance’s announcement: “More realistic images, more stable editing across 100 rounds.”
If that holds up, it fixes the most annoying thing about AI image editing today. Ask most models for a fifth or sixth change and the face shifts, the colours drift and parts you never touched quietly change. ByteDance says Seedream 5.5 keeps everything you did not ask to edit the same, round after round.
The official demo
ByteDance published its demos as video on the announcement page, not on YouTube. In this one, a portrait is edited one change at a time across 100 rounds:
And here ByteDance shows the new transparent output, cutting a subject out on its own layer:
Both files are large (about 100 MB each), so they load only when you press play. These are ByteDance’s showcase examples, not tests we ran.
What is new
ByteDance lists three main upgrades:
- Stable editing across 100 rounds. You can keep asking for changes to the same image, and ByteDance says the regions you did not edit are preserved at the pixel level, with no colour shift building up over time.
- Native transparency (RGBA). The model can output images with full or partial transparency, such as a product, a logo or a character on its own layer, ready to drop onto another background without cutting it out by hand.
- More realistic images and clearer text. ByteDance cites more natural lighting and skin, and sharper small text in both Chinese and English, for posters, packaging, slides and other layouts.
ByteDance’s own results
These figures come from ByteDance’s announcement. They are its own evaluations, not independent tests:
- On its MagicArena platform, ByteDance tested six kinds of work: e-commerce, marketing, film and TV, gaming, office productivity and fashion design. It says the usable-result rate in film and TV and in office productivity was more than double Seedream 5.0 Pro’s.
- On image-to-image realism against the previous version, ByteDance reports a preference (GSB) score of up to +55.62%.
- On text, it says Seedream 5.5 leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure, with a 0.4% rate of severe text errors and 1.7% for minor ones.
- It also claims a substantial advantage over those rivals in text-to-image realism.
ByteDance says a Pro version of the 5.5 series is coming next.
Where you can use it
BytePlus, ByteDance’s cloud business, lists Dola-Seedream-5.5 as now available on its ModelArk platform for developers. We did not find a 5.5 price on BytePlus’s pricing page yet. Reports also point to ByteDance’s consumer apps, such as Dreamina and CapCut, but we have not confirmed the rollout there.
What it means for working professionals
- Designers and marketers: long revision rounds are where image models usually fall apart. If the 100-round claim holds, you could refine one image with a client instead of regenerating it and starting again.
- E-commerce teams: transparent output means product shots and assets that drop straight into a listing or a layout.
- Anyone making text-heavy images: posters, menus and slides have been a weak spot for image models. ByteDance’s error rates are encouraging, but proofread every word before it goes out.
For how a rival handles similar jobs, see our hands-on review of ChatGPT Images 2.6, and for Google’s latest video model, our report on Gemini Omni 1.2 Flash.
We will be reviewing Seedream 5.5 as soon as possible, including the 100-round editing claim. Stay tuned!
Sources
Every story names its source. Check our work.