ByteDance launches Seedream 5.5 with claims of image stability across 100 edits
The AI image model is gradually reaching Doubao, Jimeng and Volcengine, with support for generating assets on transparent backgrounds.
ByteDance has introduced its Seedream 5.5 AI image model, saying it can preserve untouched areas of a picture through 100 consecutive edits. The company’s Seed research team announced the model, which is being introduced gradually through Chinese apps including Doubao and Jimeng, as well as its Volcengine cloud.
According to ByteDance, areas outside the requested changes remain pixel-identical, without shifts in color. A company demonstration used a portrait for 100 successive edits, changing clothing and objects while also modifying details such as rings.
The model also generates assets on transparent backgrounds, allowing designers to place them into layouts without manually removing a background. ByteDance claims improvements in the rendering of skin, lighting and tightly packed small text.
Internal tests put Seedream 5.5 ahead of OpenAI’s GPT Image 2.5 and Google’s Nano Banana 2.1 for text and realism, ByteDance said. The Decoder noted that those findings lack independent verification. ByteDance reported serious text errors in 0.4% of its test images.
Seedream and ByteDance’s Seedance video model fall under a copyright agreement the company signed with Hollywood’s studios in August. ByteDance also said a Pro edition of Seedream 5.5 is coming.
