Seedream 4.0

ByteDance's Seedream 4.0, a strong general-purpose image model with particular strength in photographic realism and Chinese-language text rendering.

Seedream 4.0 is the release where ByteDance's Seed team folded its separate generation and editing models into one. Seedream 3.0 made images and SeedEdit 3.0 changed them; 4.0 does both from a single architecture, and the team says training the two jobs together produced better results than training either alone on instruction following, image quality and aesthetic appeal. That is the structural fact everything else in the post follows from: the model takes text, an image, or any mixture, and the same weights handle text-to-image, image-to-image, single-image editing, multi-image editing and composition.

Text rendering is where ByteDance claims the clearest break from earlier generations. The team says 4.0 renders text correctly and legibly while laying out genuinely difficult content - formulas, tables, chemical structures and statistical charts - which is what makes it usable for teaching material, academic figures and infographics rather than only for pictures. It also supports editing the text afterwards and swapping fonts, so a rendered layout is a starting point rather than a finished frame.

Two capabilities are unusual enough to be worth naming. Visual-signal control is native: edge, depth and mask conditioning work inside the model, without the external control networks that pipelines normally bolt on, and a rough sketch or a set of guide lines is enough to steer a composition, which matters for pose control, architectural work and interface prototypes. And the model reasons before it draws - ByteDance shows it solving puzzles, completing crosswords and continuing comic strips, tasks that need it to hold a physical or temporal constraint rather than only to match a description.

On its own evaluations, ByteDance reports 4.0 leading on every dimension of its MagicBench human-assessment benchmark for both text-to-image generation and image editing, and taking the highest Elo rating for single-image editing on MagicArena. The team is specific about where the edge sits against other models: image texture, lighting and colour. Elsewhere in the post it notes that the generation resolution ceiling rose from 2K to 4K in this generation and that the underlying transformer runs more than ten times faster than Seedream 3.0's, through distillation, quantisation and cache reuse rather than a smaller model.

ByteDance is measured about what it built. It calls 4.0 an early-stage prototype of a general-purpose multimodal creative engine rather than a finished one, and the comparative claims in the post rest on the team's own human-scored benchmarks. Treat the reasoning demonstrations as an indication of a new capability rather than a guarantee it holds on your prompt.