Testing Seedream 5.5 for Stable Image Editing and a Talking Avatar

What happened

Baoyu tests ByteDance’s Seedream 5.5 by generating a talking cartoon avatar from 12 images and repeatedly editing a detailed tourist photo. In the avatar test, all 11 edited images met the author’s pixel-difference criteria; in the photo test, eight of 10 rounds left untouched areas pixel-identical to the prior image. The article also reports mixed-language text results, opaque glass in a transparent-background image, and a successful pose-reference test.

Read next

• More changes in today’s briefingGoogle DeepMind Introduces Gemini 4 Hydrogen for Fast Agentic Workflows

The story so far

ByteDance released Seedream 5.5, an image generation and editing model that is rolling out through Jimeng, Doubao, Doubao Work, and Volcengine Ark; developers can try it on Ark as Doubao-Seedream-5.5. ByteDance says the model can preserve unchanged image areas at the pixel level through up to 100 editing rounds and create native RGBA images with transparent backgrounds. A Pro version is planned. In his tests, Baoyu found that eight of 10 rounds of edits to a crowded tourist photo left untouched areas pixel-identical to the prior image, and all 11 edits to a talking cartoon avatar met his pixel-difference criteria. He also reported mixed results for multilingual text and opaque glass in a transparent-background image. ByteDance reports a severe text error rate of 0.4%, down from 2.8% in the previous generation, and says editing realism trails Nano Banana 2.1 by about 6%. For 2K square-image editing, the reported cost is about 0.28 yuan per image, less than half the Seedream 5.0 Pro cost. Under the API pricing model, text tokens are not charged.