Stable Multi-Turn Editing
Feed each result into the next edit while preserving unrequested regions, subject details, composition, and color more reliably across long revision chains.
Quick Start is not available for this model yet.
Start from the cost of a single sample and pick a testing budget. GPTProto rates are 15% below list price.
| Scenario | Bytedance list | OpenRouter | GPTProto | You save / mo |
|---|---|---|---|---|
| Quick tests & prototypes100 images / mo | $2.80 | $2.95 | $2.38 | −$0.42≈ $5.04 / yr |
| Product & e-commerce photos1,000 images / mo | $28.00 | $29.54 | $23.80 | −$4.20≈ $50.40 / yr |
| Ad / marketing creatives2,500 images / mo | $70.00 | $73.85 | $59.50 | −$10.50≈ $126.00 / yr |
| Social media content10,000 images / mo | $280.00 | $295.40 | $238.00 | −$42.00≈ $504.00 / yr |
Use the Seedream 5.5 API to generate a new image, revise an existing one across multiple turns, combine up to 10 references, or separate a finished design into editable layers. The model outputs at 1K, 1.5K, or 2K and can create native RGBA assets with full or partial transparency. GPTProto gives developers one account and shared balance for Seedream 5.5 and other image models. Use the model when the asset must survive several revisions, move into a transparent-background workflow, or be handed to a design tool as separate visual elements.
Stable Multi-Turn Editing
Feed each result into the next edit while preserving unrequested regions, subject details, composition, and color more reliably across long revision chains.
Seedream 5.5 is ByteDance Seed's image generation and editing model, released on November 9, 2026. The official BytePlus model ID is dola-seedream-5-5-261028. It supports text-to-image generation, single-image editing, multi-reference editing, continued multi-turn revisions, native RGBA output, and layer decomposition. It generates still images rather than video.
The main upgrade is control after the first render. A normal generation or editing request returns one image. For another revision, pass the latest result back with a new instruction that states what to change and what must remain untouched. ByteDance demonstrated a 100-round edit sequence, but that example should be treated as a capability demonstration rather than a guarantee for every prompt.
Transparent generation and layer decomposition are separate workflows. Native RGBA creates one reusable image with an alpha channel. Layer decomposition takes one existing image and returns a base image plus separate PNG layers, including their stacking order and location data. For release context and evaluation details, read What Is Seedream 5.5?.
| Specification | Verified capability |
|---|---|
| Provider | ByteDance Seed / BytePlus ModelArk |
| Official model ID | dola-seedream-5-5-261028 |
| Core tasks | Text-to-image, single-image editing, multi-reference editing, multi-turn editing, and layer decomposition |
| Reference images | Up to 10 for generation or editing; exactly one input image for layer decomposition |
| Output resolution | 1K, 1.5K, or 2K presets |
| Standard output | One image per generation or editing request |
| Output format | PNG or JPEG; transparent output requires an alpha-capable PNG workflow |
| Native transparency | Text-to-RGBA and image-to-RGBA with full or partial transparency |
| Layer output | One base image plus up to 99 layers; 2–100 returned images including the base |
| Layer metadata | Stacking order, bounding box, name, description, and optional rich-text OCR data |
| Input limits | Up to 30 MB and 36 megapixels per reference image in the upstream API |
| Goal | Input | Output | Important detail |
|---|---|---|---|
| Generate a new image | Prompt only | One image | Select PNG with a transparent background when an RGBA asset is required. |
| Edit or combine references | Prompt plus 1–10 images | One revised image | Name both the requested change and the details that must stay fixed. |
| Continue a revision chain | Previous result plus the next instruction | One new image per turn | Reuse the latest output rather than restarting from the original image. |
| Decompose a design | Exactly one image; optional split instruction | Base image plus separate layers | The total output count includes the base image and can be capped from 2 to 100. |
This distinction prevents two common integration mistakes. Multi-image editing does not produce a batch of unrelated outputs, and a transparent PNG is not the same as a decomposed design. Use RGBA output for a single cutout or overlay; use layer decomposition when individual text, subject, and background elements must be moved or edited separately.
E-commerce assets: revise product colors, labels, props, or backgrounds while keeping the core product recognizable.
Transparent design elements: generate stickers, characters, icons, lettering, and product cutouts that can be placed directly into another layout.
Campaign localization: update poster copy or visual details over several turns without rebuilding the entire composition.
Layered creative tools: split a flat image into positioned PNG layers and return OCR information for editable text elements.
Brand and character systems: combine multiple references to carry recurring subjects, materials, and visual cues into a new scene.
Change only the bottle label to matte black with the exact white text “NORTH/06.” Preserve the bottle shape, cap, reflections, camera angle, shadows, background, and every object outside the label.
State the change first, then list what must remain fixed. On the next turn, use the latest output and request one additional change rather than repeating the entire redesign.
A single white running shoe in three-quarter view, natural studio lighting, clean tread detail, soft contact shadow, no text, no border, isolated for use in an e-commerce layout.
Select transparent background and PNG output in the request settings. Describe one clearly bounded subject and any semi-transparent material that must be preserved.
Separate the headline, price badge, product, foreground leaves, cast shadow, and background into independent layers. Preserve the original canvas size and the position of every element.
Use the layer-decomposition workflow with one source image. The response can include layer names, bounding boxes, stacking order, and recognized text data for downstream editing.
| Model | Choose it when | Output and workflow advantage |
|---|---|---|
| Seedream 5.5 | Repeated edits, transparent assets, or editable design layers matter most | 1K–2K output, up to 10 references, native RGBA, and a base image plus up to 99 layers |
| Seedream 5.0 Pro | You already use its established layout, multilingual text, or decomposition workflow | 1K–2K output, up to 10 references, and a base image plus up to 16 layers |
| Nano Banana 2.1 | 4K output, conversational editing, or a larger reference set is the priority | 1K–4K output and up to 14 reference images on GPTProto |
| GPT Image 2.6 Flare | Faster drafts and rapid visual iteration are the main goal | Speed-focused generation and editing with sizes up to 4K listed on GPTProto |
| GPT Image 2.6 Sunburst | Detailed compositions and precision-focused OpenAI editing are more important | Quality-focused generation and editing with sizes up to 4K listed on GPTProto |
For Seedream 5.5 vs Seedream 5.0 Pro, choose 5.5 when you need text-to-RGBA, broader image-to-RGBA input support, richer OCR data, or more than 16 layers. Keep 5.0 Pro when an existing integration already meets the job and those additions do not change the workflow.
For Seedream 5.5 vs Nano Banana 2.1, the practical split is 2K transparent and layered assets versus 4K output and a larger reference set. For GPT Image 2.6 vs Seedream 5.5, choose Flare for quick iterations, Sunburst for detail-focused OpenAI workflows, and Seedream when native transparency or decomposable layers is the deciding requirement. There is no single winner across every image task.
Guides, comparisons, and updates related to this model.
All Articles
ByteDance's Seedance 2.5 Flash runs 2x faster at ~¥0.99/s (~$0.14/s) for 30s 720p clips—no 4K. See real trade-offs, Kling 4.0 Flash vs 2.5 specs & API code.

We tested 6 AI image API routes across 18 UI, stock photo, and product outputs. Compare image quality, text accuracy, pricing, and developer fit.

See Nano Banana 2.1 and Nano Banana Pro compared in five real ecommerce tests covering product photos, editing, packaging, multilingual text, and pricing.
Input
Output
