cantina - Research Intern

• Projects may focus on post-training and inference efficiency for video generation models, reward modeling and preference-based optimization multimodal data systems, or scalable infrastructure for video model training. The primary focus will be your core project, with opportunities to contribute to applied or product-adjacent work where relevant. • What You’ll Work On • Depending on your project, you may: • Research and develop distillation methods for large-scale diffusion and flow-based video generation models, including guidance and adversarial distillation • Explore techniques that reduce inference cost while preserving or improving generation quality • Develop reward models and preference-based optimization methods to improve aesthetics, motion quality, temporal consistency, and prompt adherence • Study how base-model behavior affects post-training outcomes and use experimental findings to inform model development • Design rigorous evaluations and conduct large-scale experiments on generative video models • Contribute to evaluation harnesses, model integrations, research tooling, or other product-adjacent projects related to your core work • Document and communicate your findings through research reports, internal presentations, demonstrations, and potential conference submissions • You may be a good fit if you • Are currently pursuing a PhD or are a final-year master’s student in computer science, machine learning, computer vision, or a related field • Have research experience in generative modeling, computer vision, multimodal learning, or video generation • Have hands on experience with diffusion models, flow-based models, model distillation, reinforcement learning, preference optimization, or related post-training techniques • Can formulate hypotheses, design controlled experiments, analyze results, and communicate conclusions clearly • Are proficient in Python and have hands-on experience with PyTorch, JAX, or another modern machine learning framework • Are comfortable working independently on an open-ended research problem while collaborating closely with a mentor and the broader team • Experience with video, image, audio, or other multimodal data is valuable. Publications at leading venues such as NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, or AAAI are a plus, but are not required. We care most about the quality of your thinking, the depth of your technical work, and your ability to learn quickly. • What You Can Expect • A defined project and named senior mentor before your first day • Weekly one-on-one meetings and clear project milestones • A meaningful compute allocation for your research