Luma AI Interview Guide (2026): Diffusion, Video Models, and the Track-Split Loop

Luma AI (Luma Labs, Inc.) is a San Francisco Bay Area company building generative models for video, images, and 3D. It was founded in 2021 by Amit Jain, who led 3D computer vision work at Apple, and Alex Yu, a UC Berkeley researcher from the early NeRF and Plenoxels work on neural 3D scene representation. The company is best known for Dream Machine, its video and image platform on the web and iOS, which has passed 30 million registered users. Its model line includes the Ray video models (Ray2 in early 2025, then Ray3, a reasoning video model with HDR output, in late 2025) and Photon for image generation, with a stated goal of building multimodal models that simulate the physical world. In November 2025 Luma raised a $900 million Series C led by HUMAIN, backed by Saudi Arabia’s Public Investment Fund, with AMD Ventures, Andreessen Horowitz, Amplify Partners, and Matrix Partners joining, at a valuation above $4 billion.

Process

Public detail on Luma’s loop is thin, so treat this as the likely shape drawn from a few Glassdoor and forum accounts plus the company’s stage, not a fixed script. Expect a recruiter screen first, then a path that splits by track. For engineering roles, plan on one or two technical rounds heavy on coding and practical builds, with ML system design for anything that touches training or serving infrastructure. For research and ML roles, expect a walkthrough of your past work plus hands-on model rounds; candidates report debugging a denoising network in a Colab notebook rather than solving abstract puzzles. Most loops close with a round with the CEO focused on experience and culture fit, and a behavioral read on how you work. Glassdoor puts the average process near three weeks, so plan for roughly two to four weeks from first call to decision.

What they actually ask

• Diffusion and video modeling: how denoising diffusion works, sampling steps and schedulers, classifier-free guidance, and the trade-offs between latent and pixel-space or temporal video architectures.
• Hands-on ML debugging: fix a broken training or denoising network in a notebook, then make the pipeline reproducible. Candidates mention a grayscale MNIST denoising task paired with an augmentation pipeline.
• Multimodal and 3D architectures: how you would condition a video model on text and images, or how NeRF-style 3D representations relate to today’s generative video work.
• Large-scale training: data pipelines, checkpoint management, mixed precision, and keeping runs stable across many GPUs, plus what tends to break when you scale compute by 10x.
• Inference optimization for video: cutting latency and cost per generated clip through batching, caching, and quantization, since serving video is far heavier than serving text.
• Strong coding and evaluation: medium-level problems solved cleanly under time with clear reasoning, plus how you would measure video quality, motion coherence, or prompt adherence.

Levels and comp (2026)

Luma is well funded after its $900 million round, so it pays like a competitive SF AI lab, but it does not publish bands and public data points are sparse, so treat any single figure with caution. Compensation is base plus equity plus benefits, with the equity almost always on a four-year vest and a one-year cliff. For a company at this stage and location, expect base salaries in a broad range that climbs with level, from roughly the high $100Ks for early-career engineering into the $200Ks and above for senior and staff, with research scientists on modeling teams often at the top of that range or beyond it. Equity is the real variable: grants are private shares whose value depends entirely on the company’s outcome, not on the headline number, so a grant sized against a $4 billion-plus valuation can be worth a lot or a little depending on how things play out. Ask the recruiter for the base, the number of equity units, the current preferred-share price, and the vesting terms so you can value the offer yourself, and check levels.fyi and recent Blind threads for current bands before you anchor on any figure.

Prep priorities

Put real time into the fundamentals of diffusion and generative video: be able to explain denoising diffusion, samplers, classifier-free guidance, and why the temporal axis in video makes training and inference expensive. Rehearse a hands-on ML debugging session in a notebook, since Luma tends to test whether you can actually fix a model and make a pipeline reproducible, not just talk about one. Keep your coding sharp on medium problems and narrate your reasoning as you go. Prepare a tight walkthrough of one or two projects where you owned real modeling or systems trade-offs, with numbers on what changed. If you are targeting an infrastructure role, review training and serving at scale, including checkpointing, distributed training, and the latency and cost of generation, and be ready to tie it to video rather than generic web services. Come in with sharp questions about the model roadmap and what shipping to 30 million-plus Dream Machine users looks like day to day.

Frequently Asked Questions

Is the Luma interview mostly LeetCode? Less than at a big company. Reports point to hands-on ML tasks such as debugging a denoising network and building a reproducible augmentation pipeline, alongside medium coding. Engineering roles see more standard coding and system design, while research roles lean toward modeling depth and a walkthrough of your past work.

How hard is it, and how long does it take? Glassdoor rates difficulty around 3.1 out of 5 with mixed reviews, and lists an average process near three weeks. Staff and research-engineer loops tend to be rated the hardest. Plan for two to four weeks and expect a real modeling bar, not just puzzles.

Do I need diffusion and video experience to get in? For research and ML roles it helps a lot, since the work is generative video, images, and 3D. For infrastructure, product, or platform roles, strong systems and coding skills plus genuine interest in the domain can carry you, but you should still understand how these models train and serve.

Is there really a round with the CEO? Multiple accounts mention a closing round with the CEO focused on experience and culture fit. Treat it as a real evaluation: know why Luma specifically, be clear about what you want to build, and be ready to talk through your hardest technical call and what you learned from it.