ByteDance’s Seed team introduced Seedream 5.5, a new image model it says can keep unedited areas stable across 100 rounds of edits and can generate native RGBA images with transparent backgrounds, the company said Nov. 9. The Beijing News reported the same launch.
Why it matters: Seed said the model is aimed at professional production, where a team may keep revising the same image. It also said Seedream 5.5 improves lighting, materials, skin texture and small-text rendering, and reduces heavy gloss and overexposure.
The company said targeted training lets the model preserve unedited regions at the pixel level so later edits do not drift in color or lose detail. It also said the model can output an alpha channel, including partial transparency, so generated elements can go into design tools without a separate background-removal step. The Beijing News said Seedream 5.5 is already in ByteDance’s Doubao, Doubao Work and Jimeng apps, with an API coming to the Volcano Engine (Ark) platform, and that a 2K square-image edit is priced at about 0.28 yuan, less than half the cost of Seedream 5.0 Pro under the same input-plus-output pricing, with text tokens not billed.
ByteDance said a Pro version of 5.5 is still planned, with support for more complex context and instructions, The Beijing News reported. The Seed blog said evaluations show a substantial improvement over the previous version and leading results in text structure and text-to-image realism compared with other major models; those comparisons were not independently audited in the announcement.