Abstract:
On November 9, the ByteDance Seed team released Seedream 5.5, a new generation image creation model, focusing on improving the stability of multiple rounds of editing, the ability to generate transparent materials, and the realistic texture.
In terms of multi-round editing, Seedream 5.5 achieves pixel-level maintenance in non-editing areas through special training, which can eliminate color shifts, reduce detail loss and image quality degradation in continuous modifications, and support up to 100 rounds of continuous editing around the same work. In the official demonstration, after 100 rounds of changes in clothing, accessories, and desktop objects, the model can still finely modify small objects such as rings and frames, and handle occlusion and light and shadow connections well.
Native RGBA transparent material generation is a new capability of this version. It can directly output transparent or semi-transparent images with Alpha channels, finely process outline edges and transparent areas, and eliminate the need for post-cutting and trimming. The generated transparent material can be reused across scenes as independent assets. It also supports integrating the strokes and colors of the reference image into the specified text to generate word art with a transparent background.
In terms of image quality and text, the model has improved the lighting, material and skin texture performance, reducing excessive gloss and overexposure; the text rendering capability has been significantly improved, and dense small Chinese and English characters can be restored more accurately. According to official evaluation, the serious text corruption rate dropped to 0.4%, and the minor text collapse rate dropped to 1.7%, both of which are better than the previous generation.
Evaluations show that compared to Seedream 5.0 Pro, 5.5 has improved generation effects and usability in six scenarios including e-commerce, marketing, film and television, games, office and clothing design. Among them, the usability rate in film and television and office scenes is more than twice that of the previous version. Compared with mainstream models such as GPT Image 2.5 Sunburst and Nano Banana 2.1, it is leading in terms of text structure and textural image realism.
Currently, Seedream 5.5 is being launched on platforms such as Jimeng, Doubao, Doubao Work and Volcano Ark. Officials stated that a Pro version will be launched in the future and the application of the model in Agent workflow will be expanded.
Comments