Senior Research Scientist Video Generation Post-Training

Job Description:

• Video Generation and Post-Training: Design and develop post-training approaches for video generation models to improve visual quality, instruction following, motion realism, and overall generation fidelity.

• Video Enhancement and Restoration: Explore generative approaches to video super-resolution, restoration, and editing, enhancing detail and perceptual quality while preserving temporal consistency.

• Data and Preference Alignment: Develop high-quality instruction and preference datasets for video post-training. Explore alignment techniques such as RLHF, DPO, GRPO, and PPO to improve generated videos according to human preferences.

• Video Quality Evaluation: Build evaluation pipelines covering visual fidelity, temporal consistency, motion quality, prompt adherence, and human preference. Research video quality assessment methods and multi-dimensional reward models to support model optimization.

• Controllable and Consistent Generation: Advance capabilities in long-video generation, storyline consistency, object and scene consistency, and precise control over camera movement and motion.

• Efficient Inference: Research and implement model distillation and other acceleration techniques for diffusion-based video models, improving inference speed and reducing computational and memory requirements.

• Research and Collaboration: Stay current with advances in video generation and multimodal AI, translate research ideas into scalable solutions, and collaborate with researchers, engineers, and product teams.

Requirements:

• Master's degree or above in Computer Science, Artificial Intelligence, Computer Vision, or a related field.

• Minimum 2 years of full-time research experience in computer vision and generative AI, with hands-on experience in video generation, video restoration, video super-resolution, or video editing.

• First-author publication(s) at top-tier conferences (CVPR/ICCV/ECCV/ICLR/NeurIPS) on video generation, video enhancement, or generative models.

• Demonstrated experience developing diffusion-based generative models for visual content generation or enhancement, with a strong understanding of visual fidelity, temporal consistency, and perceptual quality.

• Experience with model post-training or fine-tuning, and familiarity with approaches such as parameter-efficient fine-tuning, preference alignment, or model distillation.

• Experience designing video quality evaluation methods or pipelines, including the assessment of perceptual quality and temporal consistency.

• Strong programming and experimental skills, with the ability to investigate research problems, design experiments, and translate research ideas into effective solutions.

• Good problem analysis and research skills sustained curiosity about advances in video generation and multimodal AI and the ability to work independently and collaboratively across research and engineering teams.