AI News
All stories
Reported

ByteDance releases Seedream 5.5 image model with multi-turn editing

Matthew Berman calls ByteDance's Seedream 5.5 a top image generator and says it supports multi-turn editing that keeps images consistent.

Models · · 2 min read

ByteDance's Seedream 5.5 is an image generation model. Its maker says it can apply specific changes over several rounds of editing while leaving the rest of the picture alone. Matthew Berman called it "probably the best image generation model out there right now" and said its multi-turn editing keeps the image consistent. In his account, the user tells the model exactly what to change, and everything else should stay put.

ByteDance's Seed product page describes precise, targeted changes across successive edits, with unedited regions kept "intact at the pixel level." Its examples show a room where furniture and lighting are added over several edits while the perspective stays coherent. Another shows a portrait where the hairstyle and outfit change while the background holds. A third shows a watch face where the time and charts update while the layout stays the same. The page presents these as deliberate design goals. It does not explain how the model keeps untouched pixels stable from one round to the next, and it does not say how often an edit drifts.

Berman's evidence is visual rather than measured. He pointed to hair detail, skin texture, the fabrics of a woman's sweater, and the fleece on the outside of a jacket. He also pointed to depth of field: a waiter walking past gets some blur, while two diners stay completely in focus, and the background stays blurred even though nothing in it moves. He said he would probably think one image "is absolutely real." These are his judgments of images he showed on screen.

Berman's ranking also sits awkwardly with the rest of his segment. Shortly before he covered Seedream, he presented OpenAI's GPT Image 2.6 and said that "it seems like maybe" OpenAI is the only company with a good image model, or the best one. Calling Seedream probably the best places it above a model he had just praised, and he does not explain the difference.

The Seed page also claims native support for transparent and semi-transparent backgrounds, so a single prompt can produce each image as a separate layer with a transparent background. Every example on the page, including the sticker sets and the yarn-textured lettering, is ByteDance's own.

What changed

ByteDance released Seedream 5.5, an image model that applies targeted changes over several editing rounds while keeping unedited regions intact.

Type: New capability · Applies to: Seedream 5.5

Why it matters

Multi-turn consistent editing matters for product and marketing teams that need several revisions of the same image.

Original source

Where we saw it

1 source

Forward Future (Matthew Berman)

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works