Hong Li

Mixture-of-LoRAs models can finally leverage their full potential: ReMix solves the long-standing problem of imbalanced LoRA usage by using reinforcement learning to train a router with non-learnable weights, leading to significant performance gains.

Control realistic face animations like never before by disentangling identity and motion, enabling arbitrary expression interpolation in an unsupervised manner.

Generate high-fidelity 3D avatars in seconds, not minutes, by directly mapping multi-modal prompts to 3D representations using a dual diffusion model trained on a new large-scale dataset.

Data-parallel LLM fine-tuning can appear stable based on global metrics, but this paper reveals significant hidden worker-level divergence that can be diagnosed with simple metrics.