World-Model Architect: Controllable Generative Video

Luma; Redwood City (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

Luma in California seeks a researcher to turn its generative video models into world models: interactive, controllable, physically faithful, and useful for embodied reasoning. You will invent architectures (diffusion/transformer/AR hybrids), develop action/view conditioning, define metrics for fidelity and coherence, run scaling studies, and publish toward an open-source release.

This role targets PhD-level expertise in ML, CV, or robotics with strong PyTorch skills and a track record of

Qualifications

• PhD or equivalent research record in ML, CV, robotics, or related field.
• Strong expertise in large-scale generative modeling, self-supervised learning, or model-based RL.
• Proficient PyTorch and multi-node distributed training experience.
• Proven research track record with top venues or widely used releases.

Responsibilities

• Invent world-model architectures focused on controllability and physical consistency.
• Develop controllability mechanisms for action and view conditioning.
• Define metrics for physical fidelity, long-horizon coherence, and policy usefulness.
• Publish at the frontier and contribute to the open-source release.
• Run scaling studies and push promising directions toward open release.

Skills

Large-scale generative modeling

Self-supervised learning

Model-based RL

PyTorch & distributed training

Research publications

Education

PhD in ML/CV/Robotics

Job description

Luma in California seeks a researcher to turn its generative video models into world models: interactive, controllable, physically faithful, and useful for embodied reasoning. You will invent architectures (diffusion/transformer/AR hybrids), develop action/view conditioning, define metrics for fidelity and coherence, run scaling studies, and publish toward an open-source release.

This role targets PhD-level expertise in ML, CV, or robotics with strong PyTorch skills and a track record of