Leader of AIGC Video Generation Algorithm

Job Description

• Formulate the technical roadmap for AIGC including both pre-training and post-training phases, covering distributed training infrastructure, model alignment, evaluation, and deployment for inference.
• Lead the design and optimization of distributed training toolchains for very large-scale generative AI models, focusing on computation, communication, and storage efficiency while ensuring training stability.
• Oversee architecture design post-training for video generation, including curating high-quality instruction datasets, applying preference alignment techniques such as RLHF, DPO, GRPO, PPO, and enhancing video quality pipelines.
• Advance capabilities in long-duration video modeling, maintaining storyline coherence, precise camera controls, and supporting multimodal content generation.
• Develop frameworks for video quality assessment and multidimensional reward modeling to systematically evaluate and improve video output quality.
• Lead efforts in model distillation, quantization, and accelerating inference to transition research models into production environments.

Requirements

• Master’s degree or higher in Computer Science or a related discipline.
• A minimum of five years of pertinent experience in AI or machine learning with a strong emphasis on generative modeling.
• Proven leadership capabilities including managing technical teams, hiring, mentoring, and talent retention.
• Hands-on expertise in AIGC pre-training and/or post-training phases with deep knowledge of Transformer architectures and diffusion models like Stable Diffusion, Flux, or DiT.
• Strong comprehension of distributed training techniques such as data, pipeline, tensor, or expert parallelism; familiarity with frameworks like PyTorch, DeepSpeed, or Megatron-LM; alternatively, solid experience in preference alignment strategies (RLHF, DPO, GRPO, PPO), fine-tuning methods (LoRA, QLoRA, DoRA), and distillation approaches (Consistency Models, Flow Matching).
• Excellent communication skills for cross-functional collaboration and presentations to senior leadership.

Additional Advantages

• Experience in establishing new technical teams or functions from scratch.
• Complete ownership experience over the entire video generation model lifecycle, from data acquisition to deployment.
• Leadership in research related to physical simulation, world and temporal consistency, or causal inference.
• Expertise in large-scale video evaluation and human preference alignment techniques.
• Authorship in leading academic venues such as NeurIPS, ICML, ICLR, CVPR, ACL, or EMNLP.
• Knowledge of GPU hardware, CUDA programming, NCCL, and cuDNN libraries.
• Experience in optimization techniques like inference acceleration, VRAM compression, and quantization.