Video to Video
This project combines live-action performance, hand-drawn imagery and generative AI. We began by filming our main character on a low-quality camera, then Moritz painted around 40 frames in a rough, expressive black-and-white style.
These drawings became the training material for a custom Qwen LoRA, allowing us to generate new scenes and landscapes with a consistent visual language. Working in ComfyUI with Wan 2.2 VACE, we animated the scenes using the original footage as motion guidance, preserving the character’s gestures, timing, and physical presence.
The result brings the texture of handmade drawings together with the spontaneity of a recorded performance, extending a small collection of footage and paintings into a larger cinematic world
Selected Sequence
Sound on
|
Fullscreen
Video
Sound on
|
Fullscreen
Selection 40 handdrawn paintings Dataset

⬑ Process / Merlin78
[17. December 2025]
In collaboration with Moritz Mader
Helpful Resources
- ComfyUI Documentation — Node-based workflow interface for building and experimenting with AI pipelines.
- Qwen-Image-2.1
- WAN 2.2 VACE — Video generation model with Video-Aware Consistency Engine for stable, coherent frame-to-frame output.
- LoRa Training on Ostris AI Toolkit
- Youtube: Purz and Machine Delusions explain Wan2.2 Vace












