Research Scientist, Real-Time Video Generation

About Cantina

Cantina Labs is a social AI company, developing a suite of advanced real-time models that push the boundaries of expression, personality, and realism. We bring characters to life, transforming how people tell stories, connect, and create. We build and power ecosystems. Cantina, our flagship social AI platform, is just the beginning.

If you're excited about the potential AI has to shape human creativity and social interactions, join us in building the future!

About the Role

We’re working on real-time video generation for interactive characters, spanning realistic humans, stylized characters, and non-human identities. We want these characters to feel present: to speak, react, and express themselves through their face, gestures, and body language.

This brings together a set of closely connected research problems. A character needs a consistent appearance, expressive and coherent motion, and a natural relationship between speech and movement. The model also needs to respond quickly and keep generating reliably as an interaction unfolds. Quality, controllability, and efficiency all matter.

You’ll work across streaming generation, fine-tuning, distillation, inference optimization, and data. Some problems will call for a new training approach; others will come down to better examples, a more useful evaluation, or understanding a bottleneck in the implementation. Part of the job is figuring out which intervention is worth making.

You’ll have substantial freedom to propose directions, investigate unfamiliar ideas, and shape what we work on next. We’re looking for someone who enjoys that freedom and can turn it into a steady sequence of useful experiments.

What You’ll Work On

Real-time and streaming generation.

Develop models that support responsive interaction while maintaining visual quality and consistency over time. Investigate latency, continuity, and how model behavior changes during longer generation.

Character expression and motion.

Explore how appearance, facial expression, gesture, and body movement come together into a coherent character. Improve the relationship between speech and motion across different identities and visual styles.

Fine-tuning and post-training.

Adapt video models to new data, behaviors, and capabilities. Investigate training objectives and conditioning methods, and follow promising ideas through implementation and evaluation.

Distillation and inference optimization.

Make generation faster and more efficient. Work across model-level approaches and runtime improvements, measuring their effects on quality, latency, and memory.

Data development.

Spend meaningful time with the actual training examples. Identify missing coverage and recurring failure modes, improve curation and sampling, and develop data that helps models learn the behavior we want.

Evaluation.

Build comparisons that make progress visible and failures understandable. Combine quantitative measurements with careful inspection of generated video and the experience of interacting with the model.

Research and collaboration.

Read relevant work, reproduce and adapt useful ideas, and share early experiments, demos, and findings. Help the team decide what deserves more time and compute.

What You’ll Bring

• Experience training or fine-tuning video or related generative models.
• Ability to learn unfamiliar methods quickly and turn ideas into useful experiments.
• Good experimental judgment, ownership, and clear communication.
• You’ll have freedom to explore new approaches. We expect fast iteration and early sharing of results.