April 19th.ByteDanceNew Text to Image Model Seedream 3.0.Its performance excels in both internal and external reviewsIt is the first system of its kind in the world, surpassing its predecessor, Seedream 2.0, and going head-to-head with mainstream systems such as GPT-4o, Midjourney v6.1, and Imagen 3.
The model doubles the amount of training data, adds new defective images with preprocessing masks, and uses new techniques such as resolution-adaptive sampling and mixed-resolution training to ensure high-fidelity outputs for images of different sizes.
Seedream 3.0 supports native 2K (2560×1440) resolution and takes about 3 seconds to generate a 1K (1920×1080, not specified in the paper) image. In benchmark tests such as Artificial Analysis Arena, Seedream 3.0's image quality score (Arena ELO 1158) exceeds that of GPT-4o (1157), demonstrating strong competitiveness.
Seedream 3.0 shines when it comes to text-intensive tasks, rendering English and Chinese text with a success rate of up to 94%, even with complex typesetting.
The model training dataset contains detailed aesthetic and stylistic descriptions, allowing it to outperform GPT-4o in design tasks such as posters and stickers, and even rival specialized platforms such as Canva, ByteDance revealed.
In the field of realistic portraits, the model generates more realistic details such as skin texture, wrinkles and hair, avoiding the common "over-smoothing" problem of AI portraits, with better results than Midjourney v6.1, and outputting high-resolution images without post-enlargement processing.
ByteHopper has also launched a companion tool, SeedEdit, which focuses on in-image text and image editing features. SeedEdit is said to be superior to GPT-4o and Gemini 2.0 Flash in terms of precise editing, and can complete operations such as text removal, replacement or insertion without destroying the overall style of the image, and with virtually no noticeable flaws.
ByteDance plans to integrate Seedream 3.0 into its chatbot platform "Doubao" in the future to further expand application scenarios.