Article · AI translation
Thanks to IT Home readers 咩咩洋 and Domado for the tips!
IT Home reported on November 9 that ByteDance's Seed team released Seedream 5.5, a new generation image creation model, today, focusing on three core capability upgrades:
More stable multi-round editing: It preserves non-edited areas at the pixel level, eliminates color shifts, reduces the loss of detail and image quality degradation during successive changes, and supports up to 100 rounds of continued creation on the same work.
Native RGBA transparent asset generation: It directly outputs transparent or translucent images with an Alpha channel, handles contour edges and transparent areas with precision, and eliminates the need for background removal and edge cleanup afterward, allowing generated elements to enter subsequent design workflows as standalone assets.
More natural textures and clearer text: It further improves the rendering of lighting, materials and skin texture, reduces excessive glossiness and overexposure, and produces an image quality closer to real photography. It also significantly improves text rendering, reproducing complex characters and densely packed small text more accurately.
In multi-round image editing, Seedream 5.5 follows instructions more closely and preserves non-edited areas at the pixel level. In hands-on tests, with each round's result used as the input for the next, the subject's features, composition proportions, and lighting and shadow style outside the designated editing area remained consistent after more than 100 consecutive rounds of changes.
To evaluate the model's performance in production-level creative settings, the team conducted targeted tests across six scenarios: e-commerce, marketing, film and television, gaming, office work, and clothing design. The evaluations were carried out on the MagicArena platform, where senior experts in image generation evaluation scored the generated results and their usability.
Compared with Seedream 5.0 Pro, Seedream 5.5 improved both its generated-result scores and usability rates across all six scenarios. In film and television and in office work, usability rates were more than twice those of the previous version.
For the two targeted evaluations of text structure and realism, evaluation experts compared their preferences for the generated results using GSB (Good / Same / Bad) scoring, covering both text-to-image and image-to-image tasks.
Compared with the previous generation, Seedream 5.5 received positive preference scores in all four comparisons across the two dimensions and two task types, with a GSB score of +55.62% for image-to-image realism. Compared with mainstream models such as GPT Image 2.5 Sunburst and Nano Banana 2.1, it led in text structure and text-to-image realism.
The improvement in text rendering is also reflected in a lower text error rate—that is, the proportion of generated images containing incorrect or missing characters or disordered text structures. In evaluations covering Chinese and English as well as densely packed small text, Seedream 5.5's severe text error rate fell to 0.4%, and its minor text error rate fell to 1.7%. Both measures were significantly better than those of the previous version and also ranked among the best in comparisons with mainstream industry models.
Seedream 5.5 is currently being rolled out across platforms including Jimeng, Doubao, Doubao Work, and Volcano Ark. IT Home noted that, on pricing, the Seedream5.5 API retains a two-part billing structure for “input image + output image,” with no charge for text tokens. For editing a 2K square image, the cost is about 0.28 yuan per image, less than half that of Seedream 5.0 Pro.
ByteDance said a Pro version of the Seedream 5.5 series will be released later. It will support more complex contextual inputs and creative instructions in the future, continuing to improve the model's performance on professional tasks. The team will also expand the model's use in Agent workflows around actual production needs in film and television, e-commerce, advertising, and design, while continuing to explore more general-purpose visual intelligence.