Research engineer/Scientist- Post Training

About The Role

At Luma, the Post-training team is responsible for unlocking creative control in the world’s largest and most powerful pre-trained multimodal models. The team works closely with the Fundamental Research team and the Product teams across Luma to train our image and video generative models improving their capabilities in the final step refining them to be better aligned with what our users expect.

What You’ll Do

• Optimize Luma's image and video generative models through targeted fine-tuning to improve visual quality, instruction adherence, and overall performance metrics.
• Implement reinforcement learning techniques including Direct Preference Optimization and Generalized Reward Preference Optimization to align model outputs with human preferences and quality standards.
• Partner closely with the Applied Research team to identify product requirements, understand diverse use cases across Luma's platforms, and execute targeted fine-tuning initiatives to address performance gaps and enhance user-facing capabilities.
• Conduct comprehensive side-by-side evaluations comparing model performance against leading market competitors, systematically analyzing the impact of post-training techniques on downstream performance metrics and identifying areas for improvement.
• Develop advanced post-training capabilities for Luma’s video models including Camera control, Object & character Reference, Image & Video Editing, Human Performance & Motion Transfer Approaches.
• Architect data processing pipelines for large-scale video and image datasets, implementing filtering, balancing, and captioning systems to ensure training data quality across diverse content categories.
• Research and deploy cutting-edge diffusion sampling methodologies and hyperparameter optimization strategies to achieve superior performance on established visual quality benchmarks.
• Research emerging post-training methodologies in generative AI, evaluate their applicability to Luma's product ecosystem, and integrate promising techniques into our Post-training recipe.
  
  

Qualifications

• Advanced degree (Master's or PhD) in Computer Science, Artificial Intelligence, Machine Learning, or related technical discipline with concentrated study in deep learning and computer vision methodologies. Demonstrated ability to do independent research in Academic or Industry settings.
• Substantial industry experience in large-scale deep learning model training, with demonstrated expertise in at least one of Large Language Models, Vision-Language Models, Diffusion Models, or comparable generative AI architectures.
• Comprehensive technical proficiency and practical experience with leading deep learning frameworks, including advanced competency in one of PyTorch, JAX, TensorFlow, or equivalent platforms for model development and optimization.
• Strong orientation toward applied AI implementations with emphasis on translating product requirements into technical solutions, coupled with exceptional visual discrimination and dedicated focus on enhancing visual fidelity and aesthetic quality of generated content.
• Proficiency in accelerated prototyping and demonstration development for emerging features, facilitating efficient iteration cycles and comprehensive stakeholder evaluation prior to production implementation.
• Established track record of effective cross-functional teamwork, including successful partnerships with teams spanning Product, Design, Evaluation, Applied, and creative specialists.
  
  

Compensation Range: $250K - $300K

At Luma, the Post-training team is responsible for unlocking creative control in the world's largest and most powerful pre-trained multimodal models. The team works closely with the Fundamental Research team and the Product teams across Luma to train our image and video generative models improving their capabilities in the final step refining them to be better aligned with what our users expect.