SCAIL-2

Highlights

• End-to-end, skipping the pose-extraction step
• Supports character replacement and multi-character scenes
• MIT license (the most permissive)
• The official ComfyUI template already includes a character-replacement version (including int8)

Scope

• Suitable for: character replacement, virtual humans, short drama clips
• Not suitable for: text-to-video

Tips

• The reference character should ideally be framed similarly to the driving video
• With multiple characters, first make sure each character reference image is clear

Cautions

• The MIT license is permissive, but training data provenance and portrait rights still need your own confirmation

Parameters

Resolution

Driving 512p / 704p (pose and replacement work better at 704p)

License

MIT

Deployment

16GB verdict

Based on the Wan2.1 scale, the int8 template can run on 16GB; the driving video must be short.

VRAM note

Loads the reference character and driving frames at the same time, so peak VRAM is higher than pure T2V.

Suggested starting point

Use the official int8 character-replacement template with a 3-second, 704p driving video.

Comparison

The MIT license is its biggest differentiator (the only one in this batch); end-to-end generation skips pose extraction, at the cost of a thin ecosystem (3.2k downloads, and only one character-replacement template). Against Wan-Animate-2, its license and multi-character support are points in its favor.