Lead Video Foundation Model Researcher

Tencent International Service Pte. Ltd.; Singapore; On-site; SGD 180,000 - 240,000; Full time

Job summary

Lightspeed Studios is seeking a PhD-level AI researcher to design and iterate the next-generation video generation foundation architecture. You will explore long-context attention, KV Cache compression, and memory mechanisms, work on high-compression video tokenizers, and advance unified modeling for multi-resolution and multi-frame-rate videos.

You will lead video pre-training at the 10–100 billion token scale, define data mixtures and curriculum learning strategies, study scaling laws, map

Qualifications

• Ph.D. in AI-related fields with first-author papers at top-tier conferences.
• Proficient in diffusion models and autoregressive generation.
• Experience training video/image generation models from scratch; strong PyTorch and distributed training skills.
• Understanding Video VAE/Tokenizers and 3+ years of research.

Responsibilities

• Design and iterate on the next-generation video generation foundation architecture.
• Explore core technologies for long video generation, including long-context attention, KV Cache compression, and memory mechanisms.
• Research high-compression-ratio video tokenizers and unified modeling for multi-resolution and multi-frame-rate videos.
• Lead video pre-training at the 10–100 billion token scale, define data mixture and curriculum learning strategies.
• Explore Scaling Laws, define scientific scale-up paths, and continuously improve key model capabilities.
• Track industry state-of-the-art, lead comparative experiments, drive roadmap decisions, and publish academic papers.

Skills

Diffusion models

Autoregressive generation

PyTorch

Distributed training

Video VAE

Tokenizers

Research leadership

Education

Ph.D. in AI-related fields

Job description

Lightspeed Studios is seeking a PhD-level AI researcher to design and iterate the next-generation video generation foundation architecture. You will explore long-context attention, KV Cache compression, and memory mechanisms, work on high-compression video tokenizers, and advance unified modeling for multi-resolution and multi-frame-rate videos.

You will lead video pre-training at the 10–100 billion token scale, define data mixtures and curriculum learning strategies, study scaling laws, map