Subscribe

Seedream 3.0

The Seedream 3.0 report describes bilingual text-to-image with native 2K output, a claimed 4-8x inference speedup and stronger complex Chinese text rendering.

Doubled dataset via defect-aware training, mixed-resolution training, cross-modality RoPE, resolution-aware timestep sampling and a VLM-based reward model in post-training. Developers state a 4-8x speedup with image quality maintained.

Date
Tuesday 15 April 2025
Lab
ByteDance
Kind
model
Access
closed API

Figures

MeasureValueMeasured by
Inference speedup4-8x
company claim, quality maintained
company

The date is the paper's first arXiv version, posted 15 April 2025, rather than the date of the product launch on Doubao and Dreamina.

Sources

  1. arxiv.org/abs/2504.11346

This record was checked against its sources on 6 October 2026. How we check