ByteDance shows 100 edits on one image with Seedream 5.5

ByteDance introduced Seedream 5.5 on November 9, focusing its image model on repeated edits and transparent assets. The model is designed to change details such as a prop or a piece of furniture while keeping the surrounding composition in place. Comfy said on November 10 that the model was available in compatible versions of ComfyUI and on Comfy Cloud.

A designer revising a draft may already have a layout and product appearance approved. Seedream 5.5 can generate a whole image or edit an existing one, and ByteDance trained it to preserve pixels outside the area an edit targets. A series of requests can build on one composition. The model can also create assets with their own transparency for use in another design.

What ByteDance showed in 100 edits

The company demonstrated 100 consecutive edits to an image of the same person. The output of each round became the input for the next. The demo changed clothes, glasses and objects on a table, then adjusted details as small as a ring or an eyeglass frame after earlier changes had accumulated. The 100 figure describes the length of that demonstration. ByteDance did not publish a success rate across multiple 100-edit runs or a pixel-difference measurement for the sequence.

The model page also shows a room with furniture and lighting changed across successive edits. Adding a piece of furniture involves fitting it to the perspective and lighting as well as drawing the object. ByteDance uses the example to show how a room can retain its layout while individual elements change. The result of a revision depends in part on whether those surrounding details survive the new request.

ByteDance also reported changes to dense text rendering. In its evaluation covering Chinese and English, which included small type packed closely together, it reported a severe text corruption rate of 0.4% and a minor rate of 1.7%. The figures describe those evaluation tasks and languages.

Transparent output and separated layers

Seedream 5.5 can output an alpha channel, which records how transparent each pixel is alongside its color. A fully transparent area lets the layer underneath show through. In a partly transparent area, such as tinted glass, some of the underlying color remains visible. An asset made this way can sit over another background without filling its clear areas with a solid color.

Generating a single transparent asset is one workflow. Separating an existing image into elements that can be edited independently is another. Comfy lists transparent output as well as a workflow for separating layers. The first can produce a reusable item, such as a sticker. The second starts from a composed image. Pika lists the background, objects and text as elements it can separate for further editing.

Where the model is available

Comfy offers an image-editing template on its cloud service and a partner node for compatible versions of ComfyUI. The node requires a signed-in Comfy account and credits, including when someone runs the ComfyUI interface on their own hardware. Comfy lists generation credits separately from the platform subscription.

Pika also announced Seedream 5.5 in its Image Studio on November 10. Its model page says it accepts a description for a new image or an uploaded picture for an edit. Pika offers a transparent background setting and shows the credits needed for the current choices before generation. That puts the model behind Image Studio settings, while Comfy offers it as part of a workflow that can connect to other editing steps.

ByteDance said it was rolling out the model in China through Jimeng, Doubao, Doubao Work and the Volcano Engine Ark experience center. Dreamina's official page tells users to select the model when it becomes available to them. ByteDance has also announced a Pro version of the 5.5 series, with a release date still to come.