AI video background swap for a fashion lookbook with Seedance 2.5

To swap the background behind a garment in a lookbook clip, give seedance-2.5 the original clip as a video reference and describe the new set in the prompt. Seed's release post says the model's green screen editing can replace backgrounds while keeping the main subject intact. Sume's docs confirm video references are honored by the Seedance 2.x models and set the clip length at 4 to 30 seconds.

Vendor claims are from the Seedance 2.5 post, read 2026-10-01. Sume's limits are from the video generation docs.

What does Seed say green screen editing does?

The post says the model can replace backgrounds and tell entirely different stories while keeping the main subject intact, and that it renders how the subject responds to the new environment: the fluttering direction of clothes, the state of hair, gait rhythm and lighting interaction. For a lookbook that is the relevant part, since fabric movement is what gives a swapped background away. The post's example prompt references the source as "@Video 1" and lists the new backgrounds by time range.

What can I send on Sume, and for how long?

The docs describe two ways to supply images. input_references are style or content references; frame_images pin the first or last frame, and if both are present frame_images wins and the job is image-to-video. Video and audio references are honored by the Seedance 2.x models, among others. Check supported_input_references in the catalog entry before sending one.

How should I write the prompt for a swap?

Mirror the structure in Seed's example: name the reference clip as the source of the subject and wardrobe, then say what the set becomes and for which seconds. Keep one swap per clip. If the lookbook has several looks, generate one clip per look so each can be judged alone; Seedance 2.5 in 30-second clips covers the length side.

What is not promised?

The Sume docs do not describe a dedicated green screen mode. They say reference videos are accepted; how closely a given prompt follows the vendor's described behavior is something to test on a short clip first.