About the role#
This position is part of the Privacy Innovation Lab. You will work on next-generation generative foundation models, specifically focusing on diffusion-based and unified generation architectures. The role involves end-to-end training architecture design, ranging from model-parallel execution and GPU efficiency to privacy-aware training infrastructure.
What you'll need#
- Currently pursuing a PhD in Computer Science, Computer Engineering, or a related technical discipline.
- Deep understanding of Diffusion, Flow Matching, or Rectified Flow.
- Familiarity with DiT or Transformer-based architectures in generative modeling.
- Ability to debug the full pipeline from mathematical formulation through code and training to generated outputs.
- Proficiency with PyTorch and hands-on experience training large-scale models.
- Practical experience with text-to-image or text-to-video models is preferred.
- Familiarity with multimodal modeling across text, image, video, or audio is preferred.
- Research publications or open-source contributions are preferred.
What you'll do#
- Design and optimize next-generation text-to-image and text-to-video models.
- Develop and improve DiT and Flow Matching or Rectified Flow-based generative models.
- Contribute to architecture improvements for Diffusion Transformers and unified model designs.
- Work on latent space, tokenization, and conditioning mechanisms.
- Optimize training stability, convergence speed, memory usage, and compute efficiency.
- Improve generation quality and consistency.
- Address technical challenges in long-sequence, high-resolution, and video generation.
- Develop strategies for efficient attention, temporal modeling, and long-context or long-latent modeling.
- Collaborate with systems and kernel engineers to implement model designs.
- Reproduce, analyze, and advance state-of-the-art generative models.
Location & details#
- Location: San Jose, California, United States.
- Modality: On-site.
- Employment: Full-time internship.
- Term: Rolling basis.


