Seedream

Seedream is ByteDance’s image generation model family, developed by the same team behind TikTok’s internal creative tools. It uses a unified architecture that integrates image generation and editing into a single model, enabling workflows where you generate an image and then iteratively refine it — all within the same model family without switching between different systems.

Architecture

Seedream uses a unified generation-editing architecture that handles both tasks in one framework. This is different from most model families where generation and editing are separate models with different capabilities. The unified design means the model understands both “create from scratch” and “modify existing” as points on a continuum, leading to more consistent results when combining both workflows.

Each version scales the model’s capabilities while maintaining this unified approach. Seedream 4.5 scored #10 on the LM Arena global leaderboard with a performance score of 1147, placing it among the top image generators available.

Version comparison

Feature	Seedream 4.0	Seedream 4.5	Seedream 5.0 Lite
Minimum output pixels	921,600	3,686,400	3,686,400
Maximum output pixels	16,777,216	16,777,216	16,777,216
Default dimensions	2048x2048	2048x2048	2048x2048
txt2img	Yes	Yes	Yes
img2img	Up to 14 images	Up to 14 images	Up to 14 images
Seed support	No	Yes	No
Generation time	~15s	~15s	~15s

Seedream 4.0 — the baseline version with strong generation quality at 1024–4096px. Good general-purpose choice.

Seedream 4.5 — improved quality through model scaling, with a minimum output area of 3,686,400 pixels. The only version with seed support for reproducible results. Ranked among top image generators globally.

Seedream 5.0 Lite — introduces multi-image blending, accepting up to 14 input images and combining them into a single coherent output. This enables complex compositing workflows like product swaps, element transfers between images, and multi-source style mixing while preserving depth, perspective, and lighting consistency.

When to use Seedream

• High-resolution output — 4.0, 4.5, and Lite support output up to 4096x4096, with 2048x2048 as default. Good for print, large displays, and detail-heavy images
• Generate-then-edit workflows — generate with txt2img, then refine with img2img using the same model family for consistency
• Multi-image compositing — Seedream 5.0 Lite supports up to 14 reference images; Flash and Pro support up to 10. Useful for product photography compositing, combining subjects from multiple reference photos, or style blending
• Asian aesthetic content — as ByteDance’s internal model, Seedream excels at content styles popular on TikTok and Douyin

For photorealistic output with style presets and multi-image editing, FLUX.2 offers more control. For text rendering in images, Recraft V4 is the better choice.

Job types

Job type	Description	ETA
inference.partner.seedream-4.txt2img.v2	Seedream 4.0 text-to-image	~15s
inference.partner.seedream-4.img2img.v2	Seedream 4.0 image-to-image	~20s
inference.partner.seedream-4-5.txt2img.v2	Seedream 4.5 text-to-image	~15s
inference.partner.seedream-4-5.img2img.v2	Seedream 4.5 image-to-image	~20s
inference.partner.seedream-5-0.lite.txt2img.v2	Seedream 5.0 Lite text-to-image	~15s
inference.partner.seedream-5-0.lite.img2img.v2	Seedream 5.0 Lite multi-image blending	~20s

Parameters

Seedream 4.0, 4.5, and 5.0 Lite:

• prompt (required) — text description, up to 4,096 characters
• resolution — use a resolution tier instead of width and height: 4.0 supports 1K/2K/4K; 4.5 supports 2K/4K; Lite supports 2K/3K/4K
• width, height — custom output dimensions, each 1–16384 pixels; both default to 2048. Output area must meet the limits in the table, with aspect ratio between 1/16 and 16. Do not combine these fields with resolution
• sequential_image_generation — disabled (default) or auto to request related outputs
• max_images — maximum generated images, from 1 to 15; input and generated image counts together cannot exceed 15
• watermark — set to true to add an AI-generated watermark (default: false)

Seedream 4.5 only:

• seed — integer 0–2147483647 for reproducible results

Image-to-image inputs:

• images — array of 1–14 uploaded image filenames for 4.0, 4.5, and Lite
• image — single uploaded image filename, available on 4.0 and 4.5. Use either image or images

Choose JPEG, PNG, or WebP through the Accept header. Use Accept: multipart/form-data, image/png to receive all outputs from sequential generation.

Prompting tips

• Resolution matters: Seedream defaults to 2048x2048. For Seedream 4.5 and Lite, the minimum output area is 3,686,400 pixels
• Multi-image blending: when using Seedream 5.0 Lite with multiple images, describe in your prompt how the images should combine — “place the product from the first image in the setting from the second image” or “blend the styles of all reference images”
• Watermark for compliance: enable watermark: true when generating content that needs to be clearly marked as AI-generated
• Iterative refinement: generate a base image with txt2img, then use img2img with the same version to refine specific aspects — the unified architecture maintains consistency

Examples

High-resolution text-to-image:

{
  "type": "inference.partner.seedream-4-5.txt2img.v2",
  "config": {
    "prompt": "A dancing cat under moonlight, detailed fur, dramatic lighting, volumetric rays",
    "width": 2048,
    "height": 2048,
    "seed": 42
  }
}

Multi-image blending (5.0 Lite):

{
  "type": "inference.partner.seedream-5-0.lite.img2img.v2",
  "config": {
    "prompt": "Place the sneaker from the first image on the wooden surface from the second image, matching the warm lighting from the third image",
    "images": ["sneaker.jpg", "surface.jpg", "lighting-ref.jpg"],
    "width": 2048,
    "height": 2048
  }
}

Guides

Generating ImagesStep-by-step guide for generating images with Prodia.

Transforming ImagesGuide for image-to-image transformation.

Seedream 5.0 Flash and Pro

Flash and Pro are additional models; existing Lite jobs keep their model identity. Both support text-to-image, one or multiple reference images (up to 10), interactive editing through prompts, and layer decomposition from exactly one reference image. Pro additionally supports standard and fast prompt optimization modes.

Job type	Description	ETA
inference.partner.seedream-5-0.flash.txt2img.v1	Flash text-to-image	~20s
inference.partner.seedream-5-0.flash.img2img.v1	Flash editing and layers	~20s
inference.partner.seedream-5-0.pro.txt2img.v1	Pro text-to-image	~20s
inference.partner.seedream-5-0.pro.img2img.v1	Pro editing and layers	~20s

Use resolution with 1K, 1.5K, 2K, or custom widthxheight; the default is 2K. Custom resolutions must contain 921,600–4,624,220 pixels and have an aspect ratio from 1/16 to 16. Use images for an explicit reference filename array. watermark defaults to false. Select JPEG, PNG, or WebP with the Accept header. Pro’s prompt_optimization accepts standard or fast.

For img2img, choose either image or images; upload the named files as multipart inputs. A nonempty prompt is required for ordinary editing and transparent output. Layer decomposition can omit the prompt.

Set layer_decomposition to true to return a base image and up to 16 PNG layers, plus layers.json containing the upstream layer positions and names. This mode accepts 1K, 1.5K, 2K, or resolution: "auto"; custom dimensions are unavailable for layers. Set background: "transparent" for transparent output from a single input image with an alpha channel.

{
  "type": "inference.partner.seedream-5-0.pro.img2img.v1",
  "config": {
    "image": "reference.png",
    "prompt": "Separate the subject and background into layers",
    "layer_decomposition": true,
    "resolution": "auto"
  }
}

Existing 4.0, 4.5, and Lite jobs also expose sequential generation through sequential_image_generation: "auto" and max_images. Successful outputs are returned together; input and maximum output image counts must total at most 15.

See the official image generation API.

The expanded controls for existing models use partner-only .v2 job types. Published .v1 schemas remain unchanged. Newly introduced models use partner-only .v1 job types.