YIPENG WANG

Led development of agentic capabilities for MAI-Image imagegen models, securing a top 3 position in the LMArena Leaderboard, trailing only OpenAI's GPT-Image-2 and Deepmind's NanoBanana 2.

Contributed to advancing the state-of-the-art in agentic image generation, demonstrating expertise in complex AI system design and optimization.

key contributor to the model training of Marble 1 - the world's first foundation model for 3D scene generation, powered by multi-view generative model.

Directed data strategy and led real-world spatial data collection efforts and PB-level internet multimodal data sourcing, establishing a comprehensive data pyramid strategy for pre-training and post-training spatial 3D datasets.

Investigated reinforcement learning methods to improve view-consistency of the multi-view diffusion model, enhancing model performance and realism.

Served as a main contributor to Pika 1.5 and 2.0, with the latter scoring #5 in the artificial analysis leaderboard.

Developed a ray-based heterogeneous data curation pipeline that processed 5PB of video data, yielding billions of video clips for training.

Engineered a large-scale distributed training framework utilizing FSDP+CP, optimizing computational efficiency and resource utilization.

Led post-training research on camera-pose conditioned video generation and ingredients-to-video features, enhancing model control and output quality.

Created a pipeline for generating detailed 3D models from user-captured images or videos, optimized for real-time rendering on standalone VR headsets.

Utilized NeRF-based reconstruction and advanced baking techniques to achieve high-fidelity and efficient 3D model generation.

Directed pioneering research initiatives to enable object interaction within user-captured scenes, expanding immersive VR experiences.

Engineered innovative algorithms to transform 2D monocular images and videos into 3D formats, enabling the creation of 3D photos, 3D videos, and stereo videos.

Utilized NeRF-based novel-view synthesis and mesh baking & streaming techniques to enhance visual quality and delivery efficiency.