In this guide, we’ll break down a workflow that combines Florence2, WAN 2.1 VACE, and smart masking techniques to composite someone onto an AI-generated The Witcher 3 inspired body.

What This Workflow Does

Download the workflow v1 here. Without arm tracking.

Download the workflow v2 here. With arm tracking.

Link to Reddit thread here.

• The reference video provides the motion data (head turns, expressions, movement) and we extract that data to pass off to Florence2

• Florence2 detects and masks the person’s head from the source video
  • Florence2 is run in caption-to-phrase grounding mode with the prompt "head".
• Remove Background and Refine Mask
  • BiRefNet-RMBG (background remover) is applied.
  • The mask is expanded and cleaned with GrowMask and MaskComposite.

• Apply your reference photo
  • A still image from The Witcher 3 is loaded.
  • WAN VACE encodes this as the reference style and clothing.

• Combine WAN text and VACE embeds, then sit back and wait for generation
  • RTX 4090 it took about 40 minutes.
  • RTX PRO 6000 took 4 minutes.

• Save and review!

Creative Applications

• Fan Edits
  • Place yourself into anything!
• Virtual Production
  • Generate quick previews of how an actor can look in full costume without expensive shoots.
• Cosplay Visualization
  • See your cosplay concept in motion before you build it.

Final Thoughts

Whether you’re a fan editor, digital artist, or just experimenting, this is a glimpse into the future of video generation.