Human-editable AI rotoscoping for visual effects production
Generative video models for animation
Stage
Not given
Not given
Founded
Not given
Not given
Based in
AU
US
Pricing model
Paid
Not given
Pricing
Token-based SaaS model with pre-purchased credits consumed per job, scaled to computation complexity. On-premise deployment available for confidentiality-bound studios.
Not given
Key features
Converts segmentation data to editable B-spline curves
Compatible with Nuke, Silhouette, After Effects and Fusion
On-premise deployment for confidential workflows
Hosted trials available
Machine learning-based pose estimation
Pyramid-JiT (P-JiT) architecture with 32×32 token reduction
4× more aggressive token reduction than Linum v2
Converges in 11× fewer samples than Linum v2
2B-parameter model with frozen Qwen text encoder
Pixel-space diffusion with multi-resolution readouts
PixelREPA alignment with DINOv3-L
Curriculum learning over noise schedule
What makes it different
Outputs editable splines instead of static pixel masks
Keyframes placed at motion apexes like professional roto artists work
Built by the team behind the first commercial AI rotoscoping plugin
Low-cost training: full 2B model trained for just $21K in compute
Controls baked into model training rather than post-hoc fine-tuning
Pixel-space compression without separate VAE bottleneck
Apache 2.0 open release with full model weights and code
Linum