Research Lead at Pika
Video foundation models
& real-time generation.
I work on video foundation models, from large-scale pretraining to post-training and real-time generation. At Pika, I lead research across model training, alignment, and distillation.
Previously at MiniMax, working on Hailuo video and image models. Ph.D. from Shanghai Jiao Tong University.
01 / Research
Real-time video generation
Pika / Research lead
Real-time, audio-driven
video generation for AI agents.
Led PikaStream 1.0 across model design, pretraining, SFT, RLHF, and Self-Forcing distillation for real-time, audio-driven video generation.
Read the researchAudio-driven video / Model training
Led Pika’s lipsync model through pretraining, DMD distillation, and SFT/RLHF, bringing expressive, audio-driven generation to the Pika Social App.
From Pikaformance to PikaStream
Video foundation models
Foundation models / Pretraining & SFT
Led pretraining and SFT of Pika 2.2 for text-, image-, and keyframe-to-video generation. Co-led training of the Pika 2.5 foundation model.
Explore Pika 2.5Also working on scalable data pipelines, video VAEs, and production video and voice systems, including Pika Speech ↗.
Recent updates
02 / Career
2025 — Present
Research Lead Sep 2025 — Present
Leading video-model post-training, real-time audio-video generation, and agentic video systems. Research spans SFT, RLHF, distillation, data, and production deployment.
Previously Research Scientist · Mar — Sep 2025
2024 — 2025
Multimodal Algorithm Engineer
Worked on Hailuo Video-01/02 and Image-01 foundation models, diffusion alignment, subject-reference generation, and large-scale image data curation.
2016 — 2023
3D avatars and generative models at MiniMax · Efficient inference at Qualcomm · Autonomous robots and visual SLAM at Oceanbotech.
03 / Publications
My earlier research explored local features, neural scene representations, and real-time robotic perception. These publications appear under the name Shenghao Li. * Corresponding author.
An online scene representation method that handles unknown camera poses, boundary ambiguity, and observation noise such as camera motion blur, producing sharp and compact representations of scenes under a wide range of shooting conditions.
A fully differentiable correspondence estimation method guided by epipolar geometry. S2LD asymmetrically establishes correspondences with consistent sub-pixel coordinates while reducing the cost of matching; salient features are explicitly located, descriptions are conditioned on both views via attention, and correspondences are refined progressively across multiple levels to reduce re-projection error.
A quantized self-supervised local feature for indirect VSLAM that stays robust to environmental interference in robot localization, while running in real time.
A single-input multi-output model for feature extraction, together with a descriptor quantization approach that embeds features into Hamming space for efficient matching on mobile platforms.
An automatic drug box detection method that exploits geometry and color priors of boxes observed from an RGB-D camera.
An autonomous exploration and map construction method built on an incremental caching topology-grid hybrid map (TGHM).
An end-to-end container code localization and recognition system built from an efficient code detector and a sequence recognizer.
04 / Background
My doctoral research focused on 3D vision, local feature learning, and neural scene representations.
Short bio ↗2020 — 2023
Ph.D., Pattern Recognition & Intelligent Systems
Advisor: Prof. Qunfei Zhao
Weichai Power Scholarship · Top 1%
2013 — 2020
M.S., Mechanical Engineering
B.Eng., Mechanical Engineering & English
During my internship at MiniMax I helped build a pipeline for creating NeRF-based animatable avatars, driven by facial keypoints, audio, SMPL, and other signals, along with a LightStage data pipeline used to capture 500+ subjects.
A 3D pose estimation method for densely stacked complex workpieces with an RGB-D camera, and a vision-based grabbing algorithm with a line-scan camera and a multi-truss system, deployed in production.
An open-source intelligent mobile platform with mecanum wheels, equipped with an RGB-D camera, 2D LiDAR, IMU, and odometry. The package provides Gazebo simulation, 2D SLAM, visual SLAM, object tracking, navigation, and autonomous driving.