ByteDance Seed releases Seedream 4.0

Background

Image models are often built as separate tools, one for generating a picture from text and another for editing an existing one. Keeping the tasks in one model means the same system can generate, change and combine images.

Release

ByteDance Seed released Seedream 4.0 in September 2025. Its technical report was posted to arXiv on 24 September 2025.^[1][2]

One model for several tasks

Seedream 4.0 unifies text-to-image generation, image editing and multi-image composition in one model.^[2]

Resolution and speed

The model outputs images from 1K to 4K resolution. It can generate a 2K image in as little as 1.8 seconds.^[2]

Inputs and outputs

It accepts multiple reference images as input and can output several images at once.^[2] A reference image gives the model material to copy, combine or change, so the user does not have to describe everything in words.

Benchmarks

The official page showed the model's standing on the Artificial Analysis text-to-image and editing leaderboards as of 12 September 2025.^[1] It also ranked first in ByteDance's internal Elo evaluation for single-image editing.^[1] An Elo rating is a score that rises when a model wins head-to-head comparisons against others, so a first place in such a ranking means it was preferred most often in those comparisons.

Multi-image composition means combining elements from several input pictures into one new image, for example placing an object from one photo into the scene of another. A model that handles generation and editing in one system applies the same visual understanding to each of these tasks.

Significance

Seedream 4.0 combined generation, editing and composition in a single system, and its speed and resolution range are the main practical points. Its ranking results are internal or from a single leaderboard, so they describe its standing on those tests rather than across all uses.