ByteDance’s Seedream 5.5 Launched: Edit a Single Image 100 Consecutive Times, Directly Generate Transparent Assets

Dongcha Beating AI News reports that ByteDance has released its next-generation image model Seedream 5.5, which supports up to 100 rounds of continuous edits on a single image and can directly generate transparent background assets. ByteDance showcased a demo of 100 rounds of continuous editing: centered on the same character, each round builds on the previous image to alter clothing, glasses, rings, and desktop items. After repeated edits, the character retains its original features, and the model can accurately adjust small details like rings and frames. The company states the new version keeps unmodified areas pixel-consistent, avoiding past issues of increasing distortion and color bias with edits. Another new feature is native RGBA transparent asset generation: for example, when generating a snowy owl with spread wings, the result comes with a transparent background, even handling tiny gaps between feathers. These assets—including stickers, text graphics, and product materials—can be directly added to posters or design drafts, eliminating manual matting steps. Generation quality has also improved: the new version reduces issues like overly smooth skin, greasy visuals, and local overexposure in portraits, and can generate more complex Chinese text and dense small characters. According to ByteDance’s internal tests, the severe text corruption rate dropped to 0.4%, and the minor text corruption rate is 1.7%. The usability of results for film and office scenarios is over twice that of the previous Seedream 5.0 Pro. Seedream 5.5 is now rolling out on Jidream, Doubao, Doubao Work, and Volcano Engine, with a Pro version set to launch later.