ByteDance Unveils Seedream 3.0, Claims It Beats GPT-4o in Image Generation

ByteDance, the creator of TikTok, has introduced Seedream 3.0, its latest bilingual (Chinese-English) image generation foundational model, claiming it surpasses OpenAI’s GPT-4o in visual creation quality.

An upgrade to Seedream 2.0, the new model uses an expanded dataset (nearly doubled), along with techniques like mixed-resolution training, cross-modality RoPE, and resolution-aware timestep sampling. These optimisations aim to enhance visual-text alignment, scalability, and image fidelity.

According to ByteDance, Seedream 3.0 can generate up to 2K resolution images, with improvements in speed—achieving 4 to 8 times faster performance without sacrificing quality. Post-training tuning includes diverse aesthetic captions and a VLM-based reward model to boost final output appeal.

"Our core architecture design inherits from Seedream 2.0 [4], which adopts an MMDiT [3] to process the image and text tokens and capture the relationship between the two modalities. We have increased the total parameters in our base model, and introduced several improvements in Seedream 3.0, leading to enhanced scalability, generalisability, and visual-language alignment," the company said in a technical paper.

In benchmark tests (Artificial Analysis Image Arena), Seedream 3.0 ranks alongside GPT-4o and ahead of models like Imagen 3. ByteDance emphasises the model’s strengths in complex Chinese text rendering and superior typesetting—areas where GPT-4o reportedly struggles.

For image editing, ByteDance's SeedEdit—built on Seedream—also outperforms GPT-4o and Gemini 2.0 in ID preservation and prompt accuracy, though it still faces challenges in more intricate tasks.

ByteDance further critiques GPT-4o for producing images with a yellowish tint and higher noise levels, while Seedream maintains consistent quality in color, texture, and visual clarity.