Seedream 4.0

Overview

Seedream 4.0 is a single ByteDance Seed model for both text-to-image generation and image editing. ByteDance's technical report describes a diffusion transformer paired with a VAE that reduces the number of image tokens, and the model generates native images from 1K to 4K. Generation and editing were post-trained jointly, with help from a fine-tuned vision-language model.

What you send decides the job. A prompt alone produces new images. One reference plus a prompt edits or restyles that picture. With several references you can dress the person from one photo in the outfit from another, or place subjects from four photos into one scene. Raise the output count and you get a related series that keeps characters and style consistent.

What you can build

• Photo edits by instruction, such as removing a person or switching on a room's lights while the window still shows dusk.
• Text changes inside a design, like a new event name and date in the original font and colors.
• Brand material from a single logo: packaging, caps, cards and lanyards in one style.
• Old photo restoration, including colorizing and scratch repair.
• Storyboards and comic pages that keep one character across frames.
• Renders from rough drafts, for example a hand-drawn wireframe turned into a UI mockup or a floor plan into a photorealistic interior.

Inputs

• Prompt (required). Describe the image or the edit. If you want several images, say how many.
• Reference images (optional). Up to 14 files or URLs in JPG, PNG or WebP, to edit or combine.
• Size. Eight presets: 2048x2048 (1:1, default), 2304x1728 (4:3), 1728x2304 (3:4), 2560x1440 (16:9), 1440x2560 (9:16), 2496x1664 (3:2), 1664x2496 (2:3) and 3024x1296 (21:9).
• Maximum number of output images. 1 to 15, default 1. References and outputs together cannot pass 15, so 5 references leave room for at most 10 images.
• Watermark. Off by default. When on, an "AI-generated" mark sits in the lower-right corner.

Outputs

Images at the chosen size. With the maximum at 1 you get one image. Above 1, the model reads the prompt and decides how many to return, up to your maximum, so a prompt that never asks for several may still return one.

Recommended settings

• Describe the subject and what it is doing in full sentences, then add style or lighting. ByteDance's prompt guide prefers this to keyword lists, and its API reference suggests staying under about 600 English words.
• Put text that must appear in the image inside double quotes.
• For edits, name the element to change and say what stays, for example "keep its pose unchanged".
• With several references, call them Image 1, Image 2 and so on, and say what to take from each.
• For a series, write the count into the prompt ("four storyboard frames") and set the maximum at least that high.

Limitations

• Only the eight presets above, each about 4 megapixels, are offered. 4K and custom sizes are not available on this page.
• ByteDance lists Chinese and English as the supported prompt languages.

Safety & compliance

If a reference image shows a real person's face, get that person's consent before using it.