About the Team The Vision-Applied Research team focuses on applied research in Generative AI and CV/Multimodal Understanding, and delivering intelligent solutions to ByteDance products, enabling users to make and share creative content in a much easier way. The team has research groups dedicated to generative models for content creation, image generation, video synthesis, intelligent image/video editing, and virtual humans.
The team is looking for a Research Engineer / Scientist who can take initiatives in designing and implementing efficient models for large-scale generative AI, with a particular emphasis on large model distillation and compression. The candidate will work on developing methods and infrastructure for transferring capabilities from foundation models into smaller, more efficient models, enabling scalable training, optimization, and deployment. Responsibilities may include, but are not limited to, distillation frameworks, model acceleration, hardware-efficient inference, and their applications.
Responsibilities
- Develop efficient algorithms and architectures for large-scale generative and multimodal models, using techniques such as step distillation, cfg distillation, quantization, and other methods to improve model efficiency (e.g., image generation, video generation, VLM).
- Advance scalable generative modeling approaches, including diffusion and autoregressive models, with a focus on acceleration and efficiency.
Requirements
Minimum Qualifications:
- B.S. in Computer Science or related fields, or equivalent experience
- Expertise in efficient models with deep understanding of computational bottlenecks and acceleration methods.
- Proficiency in training generative AI or LLM models using widely adopted frameworks and tools such as PyTorch and JAX.
- Strong communication and collaboration skills in fast-paced environments.
Preferred Qualifications
- Ph.D. in GenAI, MLSys or equivalent experience
- Extensive research experiences in broad GenAI, MLSys, LLM areas.
- Proven experiences in at least one of the following areas: image/video generation and editing; model compression (e.g., quantization, step/cfg distillation); efficient architectures (e.g., MoE, window attention); efficient model design; or reinforcement learning training methods (e.g., RLHF, DPO, GRPO).