ByteDance’s Seedream 5.5 survives 100 edits and outputs transparent assets

ByteDance’s Seed team has introduced Seedream 5.5, an image model it says can keep editing the same picture for 100 rounds while preserving untouched areas at the pixel level. It also generates transparent RGBA images natively.

The announcement is dated November 9, 2026. It describes the model as a next-generation image creation model aimed at professional visual production, and names three upgrades: steadier multi-turn editing, native alpha-channel output, and more realistic images with clearer text.

Editing that does not drift over 100 rounds

The core problem ByteDance describes is accumulation. When one output becomes the next input, small errors pile up and cause detail loss and degradation. Seedream 5.5 uses targeted training to preserve unedited regions and remove color shifts in them, according to the company.

In its main demo, the team ran 100 successive edits on an image of one person, changing clothing, accessories and tabletop objects. ByteDance says the model still handled small items such as rings and eyeglass frames, accounting for pose, occlusion and lighting. Further demos cover floral designs across several materials and an interior that changes step by step, where perspective, occlusion and cast shadows are meant to stay coherent as a box moves around the room.

The second upgrade is native RGBA output. The model writes an alpha channel directly, with full and partial transparency, so a generated element can go into layout or compositing without separate background removal or edge cleanup. ByteDance showed a snowy owl with detailed wingtip feathers, text that borrows the brushwork of a reference image, and a set of character assets reused across different scenes. Assets can be made from text descriptions or reference images.

On realism, the company says lighting, materials and skin textures look more natural, with less gloss and overexposure. Dense small text, including Chinese characters with many strokes, is also said to be rendered more accurately.

The numbers behind the claims

ByteDance measured professional output across six scenarios: online retail, advertising, film and TV, game development, workplace documents and apparel design. The tests ran on the MagicArena platform with experienced evaluators. Against Seedream 5.0 Pro, version 5.5 improved both quality scores and usability rates in every scenario, and in film and TV and in workplace documents the share of usable results more than doubled compared with the earlier model.

In paired comparisons using GSB (Good/Same/Bad) scoring, Seedream 5.5 was positive against the previous version in all four text-structure and realism tests, reaching +55.62% for realism in image-to-image tasks. For text, the severe corruption rate fell to 0.4% and the minor rate to 1.7% in Chinese and English tests that included dense small text.

ByteDance says the model leads “GPT Image 2.5 Sunburst” and “Nano Banana 2.1” in text structure across both task types and has a substantial advantage in text-to-image realism. These are the company’s own evaluations; the page does not give the scores for rival models in text, and nothing has been independently verified.

A crowded field where transparency is the differentiator

Seedream 5.5 arrives months after Seedream 5.0 Pro, which several third-party listings date to July 2026. That model is already sold through aggregators: OpenRouter lists it from about $0.045 per image and Vercel’s AI Gateway shows $0.04. Those are reseller prices for the earlier model, not a ByteDance price list, and no figure for 5.5 appears in the announcement.

Against that backdrop, the headline comparison is with OpenAI-class and Google-class rivals, and ByteDance picked text structure and realism as the battlegrounds. The more distinctive pitch is native RGBA. Many image tools still produce flat pictures that designers must cut out by hand, so an alpha channel built in at generation time targets a real production cost. Whether the edges hold up on hair, glass and smoke is something the demos cannot settle.

The 100-round claim also needs context. It rests on a demo with one subject, and the page does not publish a metric for drift per round. Anyone who has tried iterative photo edits, as covered in our guide to ChatGPT prompts for photos, knows quality loss is the usual failure, so independent tests will matter.

No access details yet, and a Pro model to come

The announcement does not state where Seedream 5.5 can be used, in which regions, or at what price. Third-party reseller pages already list a “Seedream 5.5” API, but those pages are not ByteDance documentation, and one aggregator describes the model as coming soon, so availability should be confirmed with ByteDance or its Volcano Engine and BytePlus channels.

ByteDance says a Pro version of the Seedream 5.5 series will follow, without a date. It also plans to support more complex contextual inputs and creative instructions, and to expand the model’s use in agent workflows. For ByteDance’s wider model work this month, see our report on Cola2.