📄 Curriculum Vitae

Research intern

• Proposed Ring Forcing to address the challenge of constructing and utilizing memory in long video generation. Extended the effective history span by 30× under a fixed sequence length, achieving minute-level memory and effectively maintaining long-term consistency.

Long Video Generation, AR Diffusion

Research intern (Remote)

• Designed FourTune, the first training framework for diffusion models with 4-bit weights, activations, and gradients. Reduced VRAM usage by 2.25× and accelerated training by 2.27× on FLUX.1-dev compared to 16-bit LoRA, while supporting Qwen-Image. Achieved full-precision performance in Customization, RL, and Distillation tasks.
• Core contributor to nunchaku (3.7K Stars) and ComfyUI-nunchaku (2.8K Stars).
  This library is an inference acceleration framework for 4-bit quantized diffusion models. Integrated PuLID into the nunchaku inference framework, significantly boosting inference speed and reducing VRAM usage while maintaining identity preservation and image quality.

Diffusion Model, Quantization, Acceleration, Post-Training

Research intern

• Proposed Stand-In (CVPR 2026), a lightweight and plug-and-play identity control framework for video generation, achieving SOTA face similarity and naturalness with minimal parameter and training costs.
• Designed a novel face-swapping algorithm outperforming IP-Adapter and InstantID with a 4× inference speedup. Conducted large-scale SDXL fine-tuning for high-quality generation, successfully deploying the pipeline in WeChat Channels.

AIGC, Video Generation, IP2V, IP2I, Diffusion Model