Published: November 10, 2026
Seedream 5.5 is BytePlus's newest Seedream image model: text-to-image and editing from up to 10 reference images in one model.
It's built for realistic images and precise edits that change what the prompt asks for and keep the rest of the frame intact. Next to 5.0 Pro it adds a 3K tier and transparent backgrounds, so it's the Seedream pick for cutouts and anything that gets composited.
Variants at a Glance
Type |
Modalities |
Max Res. |
Talking Points |
|
|---|---|---|---|---|
| Variant | ||||
| 5.5 (new) | Image | T2I, Editing | 3K | Adds 3K and transparent backgrounds |
| 5.0 Pro | Image | T2I, Editing | 2K | Proven predecessor, opaque output only |
Key Features
Generation & Editing
- Text-to-Image: Realistic images from a prompt, in three quality tiers up to 3K
- Multilingual Text: Readable in-image text in many languages, not just English
- Transparent Backgrounds: Transparent for a ready-to-composite PNG, or force it opaque
- Precise Edits: Changes only what the prompt names, and holds up over repeated edits
- Multi-Reference Input: Up to 10 images per request, each taggable with @, saved characters included
Technical Capabilities
Details |
|
|---|---|
| Quality / Resolution | 1.5K, 2K (default), 3K |
| Aspect Ratios | 16:9, 9:16, 1:1, 2:3, 3:4, 4:3, 3:2, 21:9 |
| Reference Images | 1 to 10 per request; JPG, PNG, WEBP, GIF, up to 30 MB each |
| Images per Run | 1 to 4 |
| Background | Auto, Transparent, Opaque; PNG output |
Limitations
- Caps at 10 reference images; merge bigger boards into a collage first
- Eight preset aspect ratios; panoramas and thin banners need a wider render and a crop
- No mask or brush: edits are steered by the prompt alone, so name the region to change
- In-image text is strong but not flawless; check small type and non-Latin scripts before shipping
Prompting Tips
- For edits, name what stays and what changes: "keep the face and light, swap only the jacket"
- Tag each reference with @ and say what it brings: the product from one, the setting from another
- Quote rendered text and place it: "the label reads 'COLD BREW' in white condensed sans"
- For cutouts, pick Transparent and describe only the subject, not a busy scene