Seedance 2.5 is online

ByteDance's Seedance 2.5 model allows fifty references per project, offering new flexibility for composing complex shots.

This new version of ByteDance’s video model extends the duration, control, and number of usable sources within a single generation. Dreamina Seedance 2.5 can produce a sequence of up to 30 seconds in standard mode, with a long option currently in testing that can reach up to three minutes. Outputs are advertised at up to 4K.

The model accepts up to fifty multimodal references. A project can thus combine a prompt, a script, images, videos, audio, visual direction, a storyboard, or simplified 3D previews. The R2V mode notably relies on green screen footage or blocking animations to more precisely guide movements, the camera, and interactions between multiple characters.

Local editing allows for correcting a portion of a sequence without having to intentionally redo the entire shot. The user can replace an object, adjust a character, or modify a background while preserving the motion, lighting, and composition already achieved. Temporal cues can also specify exactly when an action or modification should occur.

Dreamina also highlights better continuity in long scenes and multi-character sequences, as well as more precise control over shadows, camera movements, and sound direction. Dedicated workflows allow the use of renders, camera paths, or animations prepared in Blender and Maya as references, though a native plugin for these software programs is not clearly documented.

Access is through the Dreamina platform linked to CapCut. However, the model's availability seems to be rolling out gradually depending on accounts and regions: several pages already list it as accessible, while the main page still displays "coming soon."

Editorial Caveat: the first observed renders exhibit a sort of noisy, cloudy haze that is particularly unsightly. This visual flaw tends to smooth out the image and degrade overall sharpness, even on some highly polished demonstrations.