You'll own the last mile of Luma's video foundation models: making them expressive, controllable, and personalized enough for the most demanding creative work. As an Applied Research Scientist / Engineer, you sit between research, product, and our creative partners, turning state-of-the-art models into something people actually ship with.

This is a fullstack applied research role, so you'll move across modeling, data, systems, and evaluation rather than going deep in only one. The problems are specific and messy: a partner's fidelity target, an identity that has to hold across a scene, a control that has to behave. It suits someone who treats users as collaborators and cares more about real output quality than public benchmark numbers. If you'd rather optimize a single metric in isolation, this won't be a fit.

What You'll Own

• Build and maintain model variants for specific user environments and creative partners, using SFT, RL, personalization, distillation, and control adapters.

• Architect the data engine for rapid adaptation, using proprietary vertical datasets to create specialized finetunes and sharpen future training recipes.

• Define and drive end-user quality: set the success metrics, build user-aligned evaluations, and run the model/data/eval loop to hit fidelity and reliability targets in enterprise verticals.

• Partner with Product, Research, and Design to turn creative intent and user feedback into real model behavior and production-ready controls.

• Close the gap between research prototypes and production systems so the work reaches users, not just papers.

First 90 Days

One way the first 90 could unfold.

• Days 1–30 — Immerse & Diagnose: Get deep on the current models and where they fall short on controllability and personalization for priority partners, and pick the first last-mile problem worth solving.

• Days 30–60 — Ship & Validate: Deliver a model variant or control that measurably improves output for a real creative workflow, backed by an evaluation that proves it.

• Days 60–90 — Scale & Systemize: Turn that into a repeatable adaptation and evaluation loop other verticals can reuse.

What You Bring

• Strong ML fundamentals and deep experience with visual generative models (diffusion, transformers, or related architectures).

• Depth in at least one of: fine-tuning, personalization, domain adaptation, data curation, targeted distillation, interpretability, or human-feedback refinement.

• Hands-on Python and deep-learning engineering, ideally PyTorch, comfortable across prototypes and production.

• A product instinct: you treat end users and partners as collaborators and solve for their real problems.

Nice to Have

• Contributions to state-of-the-art image or video generation models.

• Experience working with creative partners (VFX, animation, film, design tools).

• A track record building workflows or tools that speed up iteration and tighten evaluation.

• Familiarity with large-scale training infrastructure and distributed systems (Ray, Slurm, Kubernetes).

About Luma: Luma's mission is to build unified general intelligence that can generate, understand, and operate in the physical world. We believe multimodality is critical for intelligence — the next step beyond language models comes from vision. Luma is an equal opportunity employer.