The_frizzy1 · Video character swap

SCAIL-2 Character Swap 12 GB (v1.1.0)

This is a ComfyUI workflow by The_frizzy1 that swaps one person in a video for a user-supplied character (a person, an animal or a cartoon) using SCAIL-2, which is based on Wan 2.1 14B (int8), together with SAM 3.1. It was submitted to the ComfyUI #OpenSourceHiggsfield challenge. It handles clips of any length, keeps the original sound, opens in App mode and needs no custom nodes. SAM 3.1 finds up to 4 matching people, sorted left to right, so the user can pick one or swap everyone for the same character, with one character per run. SCAIL-2 generates 81 frames at a time inside ComfyUI's Start Loop / End Loop. Each round starts 76 frames later, reuses the last 5 frames as overlap and colour-matches to the previous round. On an RTX 3060 12 GB (32 GB RAM, ComfyUI 0.43.1, Windows 11), a 9.4 s dance clip at 512×896 with sound kept took 24:09. Seven model files totalling 28.6 GB are required. The workflow is MIT-licensed, as is SCAIL-2 (by zai-org). Two characters per run and a face-fix pass were left out as unreliable.