Artist trains ML on real human motion, finding movement easier to learn than imitating images
pixlpa · x · 2026-10-03
Artist pixlpa shares an experimental creative pipeline: training a machine learning model on a corpus of natural human movement to generate a realistic motion signal for image work.
- The author says results exceeded expectations, with a counterintuitive observation: learning to move is a much easier problem than imitating an image
- The key trick is representing motion within a frame as an image alongside actual images
- This lets them infuse abstract imagery with an illusion of life and insert a "ghost of human intervention" into digital systems
Related event: Digital Artist Uses ML and Optical Flow to Bring Abstract Images to Life(2 posts)→
More from Multimodal
- Blogger ranks Gemini 4 Argon video below GPT-6 Astra and Opus 5.5 in early comparison — dotey · 2026-10-03
- 44-second spec ad makes Eleven v4 read a list of things AI voices 'can't do' — aloncarmel · 2026-10-03
- Opus 5.5 one-shot animates all human history as a 24-hour day, with original score — jasteinerman · 2026-10-03
- One prompt, 74 seconds: AI short film spans 30,000 years of art history — schwentker · 2026-10-03
- Free Motion2Video workflow: one image plus a dance video animates any character — EzkaProductions · 2026-10-03
- First Comprehensive Survey of Post-Training and Alignment for Video Generation Models — Arizona-State-University · 2026-10-03