Skip to main content

Introduction

Seedream is the ByteDance image model family for text-to-image, image editing, interactive editing, multi-reference fusion, and layer decomposition. These documents cover Seedream 5.5 Pro and Seedream 5.0 Pro, with separate scenario guides and copyable requests. Call synchronously through the GravitexAI gateway:
  • Base URL: https://api.gravitex.ai
  • Authorization: Bearer sk-xxxxxxxxxx
  • Content-Type: application/json

Documented models

Choose 5.5 Pro for text-only transparent output, output-count configuration, or rich-text OCR. 5.0 Pro transparent editing requires one alpha-channel reference; 5.5 Pro accepts up to 10 references without requiring alpha. Default output formats are jpeg for 5.0 Pro and png for 5.5 Pro. Check the model-specific defaults when switching.

Find examples by scenario

Supported endpoints

Both routes forward to the same upstream endpoint. These guides use /v1/images/generations for both scenarios: omit image for text-to-image, or supply a URL/Base64 string or array for image-to-image. Ordinary generation supports at most 10 references; decomposition requires exactly one.

Responses and pricing

Responses synchronously return data and usage. Default response_format is url: download data[].url within 24 hours. With b64_json, decode data[].b64_json. Ordinary generation returns one image. Decomposition returns a base image and layers with stacking order and bounding boxes; 5.5 Pro can also return rich-text OCR for regular printed text when layer_decomposition_options.text_recognition is enabled. References and successful outputs are billed separately. Each output, including the base image and individual layers, uses its actual pixel tier. Reference, generation, and decomposition prices differ between models; see each model pricing guide.

View current models and prices

The model marketplace is authoritative for current availability, capabilities, and prices.

Integration steps

  1. Create an API Key and configure the gateway URL and Bearer authentication.
  2. Choose a model and open its generation or editing guide.
  3. Copy the full JSON body. Generation requires model and prompt; editing also requires image. For decomposition, use layer_decomposition: true with one image; prompt is optional.
  4. Send a synchronous request, save the URL result or decode Base64, and inspect usage. Reconstruct layers using z_index and bounding_box.