Qiran Hu

UIUC Computer Vision and Machine Learning Group

Undergraduate Research Assistant - Advised by Professor Yaoyao Liu, 2025.05-Present, Champaign, IL

• Architect adaptive conditioning methods for 3D-aware synthetic data generation to enhance world model understanding and embodied agent performance in interactive simulations with geometry-conditioned diffusion approaches, reducing FID by 32.8% and increasing pose accuracy by 4.2x on PASCAL3D+.
• Conduct large-scale foundation model training across TB-level datasets on the National Center for Supercomputing Applications (NCSA) HPC clusters to enforce multi-view consistency with 4-bit NF4 quantization and low-level custom kernels, improving pose accuracy by 11.3% on PASCAL3D+ and reducing generation latency by 78.7% at p95.
• Design camera-controlled novel view synthesis on video generation pipelines to improve real-time perception for SLAM, visual odometry, and 3D reconstruction, reducing LPIPS by 21.0% on GSO and FV4D by 52.0% on OmniObject3D.

Links: Lab

Multimodal Continual Learning Project

Undergraduate Research Assistant, 2026.02-Present, Champaign, IL

• Selected for the NVIDIA Academic Grant Program Award to improve multimodal foundation models in class-incremental learning across audio, image, and text without catastrophic forgetting and cross-modal alignment drift, increasing R@1 by 27.6% on AudioSet.
• Advance post-hoc tensor-level weight-interpolation methods to improve multimodal retrieval through National Artificial Intelligence Research Resource (NAIRR) HPC clusters, reducing trainable parameters from 182M to 499 sigmoid-parameterized coefficients and increasing R@1 by 33.5% on AudioSet.
• Improve checkpoint fusion pipelines that merge separately trained checkpoints into a single model with no additional inference time, increasing last-task accuracy by 40.9% on UrbanSound8K.

Links: NVIDIA Grant

Long-Form Video-Language and Audio-Visual Social Understanding

Undergraduate Research Assistant, University of Illinois Urbana-Champaign, 2025.12-2026.05, Champaign, IL

• Trained streaming video-language models with temporal transformer blocks for long-form video understanding beyond 30-minute sequences, increasing zero-shot accuracy by 17.4%.
• Designed context fluidity pipelines that fused facial action units, body pose, and prosody through cross-modal attention to infer social intent and conversational role from raw recordings, achieving 89.1% accuracy on speaker-role classification.
• Built video annotation pipelines for long clinical sessions with automatic transcription and detailed labeling, achieving 90.0% accuracy on role-inversion recovery.