On November 9, Bytedance officially released the beanbag image creation model Seedream 5.5. Compared with the previous generation Seedream 5.0 Pro, the new model generation speed has been greatly improved, and the stability of multi-round editing has been enhanced. It natively supports transparent material generation, and has made key optimizations for image realism and small text rendering, which better meets the quality, stability and efficiency requirements of professional creation scenes.

Currently, Doubao, Doubao Work, and Jimeng have been synchronized to Seedream 5.5, and the API will be launched on Volcano Ark soon. In terms of pricing, Seedream 5.5 API continues the two-stage billing structure of "input image + output image", and text tokens do not incur fees. Taking 2K square image editing as an example, the cost per image is about 0.28 yuan, less than half of Seedream 5.0 Pro.

In terms of multiple rounds of image editing, Seedream 5.5 has stronger command following capabilities and achieves pixel-level maintenance of non-edited areas. In the actual measurement, the model used the results of each round as the input for the next round. After completing hundreds of rounds of modifications, the subject characteristics, composition proportions, and light and shadow styles outside the designated editing area were still consistent.

After 100 rounds of modifications, pixel-level maintenance is achieved in non-edited areas.

Seedream 5.5 natively supports the generation and splitting of RGBA transparent materials, can directly output images with Alpha channels, and supports layered output of materials. In addition to Vincentian drawings, creators can also generate the required transparent materials through reference drawings. While retaining the rich form and texture of the subject, the model will accurately handle outline edges and transparent areas, making it reusable across scenes as design assets to improve production efficiency.

Seedream 5.5 natively supports transparent material generation

In addition, Seedream 5.5 improves the performance of lighting, materials, and skin textures in realistic scenes, reducing excessive shine and overexposure, and producing results closer to natural shooting. In scenes such as backlit portraits, special effects synthesis, and multi-image fusion, the spatial relationship, occlusion levels, and light and shadow transitions between the new elements and the original image are more reasonable.

Generates results closer to natural shots

The small font rendering effect of Seedream 5.5 has also been improved, especially the Chinese generation under smaller font sizes. The structure of Chinese characters with complex strokes is still clearly legible, and can be used for information graphics, knowledge illustrations and other content that require high text accuracy.

Seedream 5.5 is fully integrated into professional-grade visual creation workflows. Based on the special test results of the MagicArena platform, the usability rate of Seedream 5.5 in six scenarios including e-commerce, marketing, film and television, games, office and clothing design is significantly improved compared to Seedream 5.0 Pro.

Source: ByteDance Seed official website

In e-commerce marketing, users can use the multi-round editing capabilities of the model to produce multiple outfit effects of the same model at a lower cost; in home decoration design, designers can quickly verify different decoration plans while keeping the spatial layout and details unchanged; in film and television storyboards and game asset production, creators can also continuously adjust a single picture without losing image quality.

ByteDance said that the Seedream 5.5 series will also launch a Pro version in the future, which will support more complex contextual input and creative instructions in the future to continue to improve the model's performance in professional tasks. At the same time, the team will focus on actual production needs in film and television, e-commerce, advertising, design and other fields, expand the application of the model in Agent workflow, and continue to explore more general visual intelligence.

Leifeng.com (Official account: Leifeng.com) Leifeng.com

edit

Resend Email Modify Email