An image creation model from ByteDance Seed with improved multi-turn editing, realism, and text rendering, plus native transparency and layer decomposition for commerce, advertising, and visual assets.
Bytedance
Released by ByteDance Seed on November 9, 2026, Seedream 5.5 focuses on stability across successive edits, realistic imagery, and clearer text. It suits iterative design work and the creation of reusable visual assets.
The model supports text-to-image creation, reference-guided generation, and editing, with expanded transparency and layer capabilities. Creators can refine character styling, product settings, or design elements before arranging the resulting assets in design tools.
| Feature | Value |
|---|---|
| Multi-turn consistency | Reduces color drift, lost detail, and degradation during repeated edits. |
| Alpha-channel generation | Supports compositing assets onto different backgrounds. |
| Multiple references | Combines subject, style, and scene guidance. |
| Layered outputs | Supports separate adjustments and recomposition. |
| Rich-text OCR information | The official layer feature can return regular text content, position, and approximate styling. |
| More natural appearance | Reduces excessive gloss and overexposure in realistic imagery. |
It supports both, with iterative editing a key focus of this release. For detailed work, establish the composition first, then refine subjects, backgrounds, and text in separate turns while identifying what must remain unchanged.
It describes a capability demonstrated in the official announcement, not a guarantee that every image remains unchanged through 100 revisions. Save milestones and compare untouched regions, colors, and details throughout the process.
No: native RGBA includes an alpha channel representing transparency. Test the asset against both dark and light backgrounds to inspect edges, translucent areas, and unwanted background residue.
Use the previous result as the next input and specify both the desired change and what must remain intact. Volcengine Ark recommends opaque RGB images for successive edits; configure parameters according to the platform’s actual support.
Open its LinkModel page and follow the published model identifier, authentication method, and request format. Validate simple generation first, then configure references, output size, and editing options according to the documentation.
Test real assets for subject consistency, text accuracy, and iterative editing quality. For transparency or layer features enabled on the platform, also inspect alpha channels, layer completeness, and result retrieval, recording configurations that work reliably.
fal charges input images per generated output. For layer decomposition, the base image determines the rate for all returned layers. Savings are shown only for matching output specifications.
| Usage | LinkModel Price | fal pricing |
|---|---|---|
| Input image | $0.006400 / image | $0.012 / input / output |
| Output image · 921,600–2,601,123 pixels | $0.016000 / image | $0.030000 / image |
| Output image · 2,601,124–2,610,000 pixels | $0.028000 / image | $0.030000 / image |
| Output image · 2,610,001–4,624,220 pixels | $0.028000 / image | $0.040500 / image |
| Layer, including background · below 921,600 pixels | $0.016000 / image | Not directly comparable |
| Layer, including background · 921,600–2,601,123 pixels | $0.016000 / image | Not directly comparable |
| Layer, including background · 2,601,124–4,624,220 pixels | $0.028000 / image | Not directly comparable |