ByteDance / text-to-image-seedream-v3
text-to-image-seedream-v3
Text-to-image seedream v3 model from ByteDance
Seedream 3.0 is a bilingual image generation model that works in both Chinese and English. It improves on Seedream 2.0 by handling complex prompts more accurately and producing better visual quality. The model generates clearer text in images, supports detailed typography, and can produce higher resolution outputs up to 2K.
To reach this level of quality, Seedream 3.0 was trained on a much larger and cleaner dataset. The team doubled the data size and used smarter sampling methods to improve variety and reduce defects. The training process also includes techniques like mixed resolution training, alignment across text and image features, and better handling of image resolution during generation. These upgrades help the model create images that better match human preferences.
Seedream 3.0 is also faster. A new acceleration method lets it generate images four to eight times quicker while keeping the same level of quality. It performs especially well with Chinese text rendering and detailed typography, which are important for many professional and creative tasks. With its strong language support and high resolution output, Seedream 3.0 offers a reliable tool for generating accurate and visually appealing images.
Example prompts
Great starting points for text-to-image-seedream-v3.
API quick start
Run text-to-image-seedream-v3 with a single API call.
{
"prompt": "A tranquil mountain lake at dawn, mist ho…",
"size": "1024x1024",
"watermark": "false",
"guidanceScale": 1
}