More control over edits, transparency, and text with Seedream 5.5

ByteDance shows Seedream 5.5 making 100 successive edits to a single image. The model also generates assets with transparent backgrounds.

Swapping out a ring or a pair of eyeglass frames in a portrait might sound simple. It gets harder when the same image has already been used to change the outfit, the accessories, and the objects on the table. In its Seedream 5.5 demonstration, ByteDance makes 100 successive edits to an image of the same person. Each request starts with the previous version. As the portrait evolves, the model continues to make small changes that account for the subject’s pose and the lighting in the scene. What needs to stay consistent from one edit to the next

A local edit can change other parts of an image. If the background shifts slightly in color during the first edit, that shift carries over into the file used for the next one. The same goes for a texture that loses definition or a detail that disappears. ByteDance says it trained Seedream 5.5 to better preserve areas outside the requested edit. Across a series of images, that matters as much for the elements that need to look the same as it does for the objects being replaced.

The previous version, Seedream 5.0 Pro, already let users target an edit by selecting a point or drawing around an area with a lasso. The 5.5 demo shows what happens when those changes are made in sequence. In one interior scene, the team moves a box. The model has to adjust the shadow, account for perspective at the box’s new position, and fill in the part of the room it previously obscured. The edit extends to the parts of the scene affected by the move. The 100-edit sequence shows how that works on one image. How often the model maintains that level of consistency on other images remains an open question. From a transparent background to a finished layout

The Seedream 5.5 model page shows two ways to get an isolated asset. The model can generate an element directly on a transparent background, such as lettering made of yarn, or extract a subject from an existing image, such as a flamingo. In both cases, the RGBA output adds opacity information to the three color channels. This alpha channel, supported by the PNG format, determines where the background shows through and how much of it is visible.

Partial transparency matters along edges and in materials that let light through. Loose fibers around a letter or the outline of a translucent object can have different levels of opacity, helping them blend into a new setting. An asset exported this way can be repositioned in a layout and reused on different backgrounds. If the lettering and subject need to move independently, each needs its own export. The model page illustrates this with a set of characters generated as separate images. What ByteDance measured in text rendering and realism

To illustrate small-text rendering, ByteDance published an English-language newspaper-style page about Shakespeare and a Chinese-language infographic about a butterfly. In its tests in both languages, which included small, tightly spaced text, the team reports a severe error rate of 0.4% and a minor error rate of 1.7%. It also evaluated text layout by asking reviewers to compare images in pairs. A page can have a clear visual hierarchy and still contain a distorted letter in a caption. The two measures capture different problems.

For realism, the examples focus on skin, fabric, and reflections. In a race scene in the rain, the car body and the wet track need to reflect light differently, while spray obscures part of the background. In a separate evaluation on MagicArena, ByteDance asked reviewers to judge image quality and usability across six production areas. In film and television, as well as office productivity, the team reports more than double the share of images rated usable compared with Seedream 5.0 Pro. Those ratings reflect reviewers’ judgments of the images they saw. Measuring how much time teams spend fixing those images in a real project would require a separate test. Where to find Seedream 5.5

In its image generation and editing guide, Dreamina lists the model as “Image 5.5 by Seedream 5.5.” The platform says access, export options, and generation costs depend on the account. ByteDance also announced a Pro version of Seedream 5.5, with a release date still to come.