ByteDance Releases Image Creation Model Seedream 5.5
ByteDance's Seed team has released its next-generation image creation model, Seedream 5.5, which supports up to 100 rounds of continuous editing around the same work and achieves pixel-level preservation of non-edited areas. The model can directly output transparent or semi-transparent images with an Alpha channel, and its severe text corruption rate has dropped to 0.4% and its minor text corruption rate to 1.7%. Seedream 5.5 is gradually rolling out on Jimeng, Doubao, Doubao Work, and Volcano Ark, with the cost of editing a single 2K square image at about 0.28 yuan, less than half that of Seedream 5.0 Pro.
Why read this
SelectedAI score 82
The original text provides specific data on multi-round editing, transparent asset generation, and text corruption rates, which can be used to assess its position in design production workflows.
Topics
Body · AI translation
One image can be edited 100 times, with transparent backgrounds output directly: ByteDance releases its new-generation image creation model Seedream 5.5
English text is being prepared.
English text is being prepared.
English text is being prepared.
More stable multi-round editing: it achieves pixel-level preservation of non-edited regions, eliminates color shifts, and reduces detail loss and image quality degradation during continuous modifications, supporting up to 100 rounds of sustained creation around the same work.
Native RGBA transparent asset generation: it directly outputs transparent or semi-transparent images with an Alpha channel, finely handling contour edges and transparent areas, eliminating the need for post-processing cutouts and edge cleanup, so that generated elements can enter subsequent design workflows as independent assets.
More natural texture and clearer text: it further improves lighting, material and skin texture rendering, reducing excessive glossiness and overexposure to more closely match the imaging quality of real photography. At the same time, it significantly enhances text rendering capability, reproducing complex glyphs and dense small text more accurately.
In multi-round image editing, Seedream 5.5 has stronger instruction-following capability and achieves pixel-level preservation of non-edited regions. In actual tests, with the model using the result of each round as the input for the next, after more than a hundred consecutive rounds of modification, the subject features, composition proportions and lighting style outside the specified edit regions remained consistent.
▲ After 100 rounds of modification, non-edited regions achieve pixel-level preservation
To evaluate the model's performance in production-grade creation scenarios, the team conducted dedicated tests across six categories of scenarios: e-commerce, marketing, film and television, gaming, office and fashion design. The evaluation was carried out on the MagicArena platform, with senior evaluation experts in the image generation field scoring the generation results and usability.
Compared with Seedream 5.0 Pro, Seedream 5.5 improved both generation quality scores and usability rates in the above six scenario categories. Among them, the usability rates in the film and television and office scenarios both reached more than twice that of the previous version.
Seedream 5.5's performance in different production scenarios. Usability rate on the left, average subjective score (MOS) on the right
The two dedicated evaluations of text structure and realism both had evaluation experts make preference comparisons of the generation results, scored using GSB (Good / Same / Bad), covering both text-to-image and image-to-image tasks.
Compared with the previous generation, Seedream 5.5 showed positive preference in all four comparisons across the two dimensions and two task types, with the GSB for realism in image-to-image reaching +55.62%. Compared with mainstream models such as GPT Image 2.5 Sunburst and Nano Banana 2.1, it leads in text structure and text-to-image realism.
Dedicated evaluations of text structure and realism, showing Seedream 5.5's relative GSB scores on text-to-image and image-to-image tasks respectively
The progress in text rendering is also reflected in a lower text corruption rate — that is, the proportion of generated images with wrong characters, missing characters or disordered structure. In evaluations covering Chinese and English as well as dense small text, Seedream 5.5's severe text corruption rate dropped to 0.4%, and its minor text corruption rate dropped to 1.7%. Both metrics are significantly better than the previous version and are also leading in horizontal comparisons with mainstream industry models.
Text rendering quality evaluation, with the severe text corruption rate on the left and the minor text corruption rate on the right; lower values indicate fewer problems
English text is being prepared.
ByteDance said that the Seedream 5.5 series will subsequently launch a Pro version, which will support more complex context inputs and creation instructions in the future, continuously improving the model's performance in professional tasks. At the same time, the team will expand the model's application in Agent workflows around the actual production needs of fields such as film and television, e-commerce, advertising and design, and continue to explore more general visual intelligence.
AI Trends