Han Cai
hcai.hm [at] gmail (dot) com

I am a research scientist at NVIDIA Research. Before joining NVIDIA, I received my doctoral degree at MIT EECS. I received my master's and bachelor's degrees at Shanghai Jiao Tong University (SJTU).

My research interests lie in artificial intelligence and machine learning, particularly efficient foundation models (diffusion models, LLMs, etc), EdgeAI and AutoML.

Email  /  Google Scholar  /  GitHub  /  Twitter  /  Linkedin



News

Selected Projects

DC-VideoGen: Efficient Video Generation with Deep Compression Video Autoencoder
DC-VideoGen is a post-training acceleration framework that can be applied to any pre-trained video diffusion model, improving efficiency by adapting it to a deep compression latent space. The framework builds on two key innovations: (i) a Deep Compression Video Autoencoder that achieves 32x/64x spatial and 4x temporal compression; and (ii) AE-Adapt-V, a robust adaptation strategy that enables rapid and stable transfer of pre-trained models into the new latent space. Adapting the Wan-2.1-14B model with DC-VideoGen requires only 10 H100 GPU days. The accelerated models achieve up to 14.8x lower latency than their base counterparts without compromising quality, and further enable 4K video generation on a single GPU.


DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space
DC-Gen is a new acceleration framework for diffusion models. DC-Gen works with any pre-trained diffusion model, boosting efficiency by transferring it into a deeply compressed latent space with lightweight post-training. For example, applying DC-Gen to FLUX.1-Krea-12B takes just 40 H100 GPU days. The resulting DC-Gen-FLUX delivers the same quality as the base model while achieving dramatic gains—53× faster inference on H100 at 4K resolution. And when paired with NVFP4, DC-Gen-FLUX (20 sampling steps) generates a 4K image in only 3.5 seconds on a single NVIDIA 5090 GPU.


Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search
Jet-Nemotron is a new family of hybrid-architecture language models that surpass state-of-the-art open-source full-attention language models such as Qwen3, Qwen2.5, Gemma3, and Llama3.2, while achieving significant efficiency gains—up to 53.6× speedup in generation throughput on H100 GPUs (256K context length, maximum batch size).


Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models
DC-AE is a new autoencoder family for accelerating high-resolution image generation. DC-AE solved the reconstruction accuracy drop issue of high spatial-compression autoencoders, improving the spatial compression ratio up to 128 while maintaining the reconstruction quality. It dramatically reduces the token number of the latent space, delivering significant training and inference speedup for latent diffusion models without any performance drop.


SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers
SANA is a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep Compression Autoencoder (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with caption labeling and selection to accelerate convergence.


Condition-Aware Neural Network for Controlled Image Generation
CAN is a new method for adding control to image generative models. In parallel to prior conditional control methods, CAN controls the image generation process by dynamically manipulating the weight of the neural network. CAN combined with EfficientViT (CaT) achieves 2.78 FID on ImageNet 512x512, surpassing DiT-XL/2 while requiring 52x fewer MACs per sampling step.


EfficientViT-SAM: Accelerated Segment Anything Model Without Performance Loss
EfficientViT-SAM is a new family of accelerated segment anything models. We replace SAM's heavy image encoder with EfficientViT. Benefiting from EfficientViT's efficiency and capacity, EfficientViT-SAM delivers 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing performance.


EfficientViT: Multi-Scale Linear Attention for High-Resolution Dense Prediction
EfficientViT is a new family of vision models for high-resolution dense prediction. It achieves global receptive field and multi-scale learning with only hardware-efficient operations. EfficientViT delivers remarkable performance gains over previous models with speedup on diverse hardware platforms. [Media: MIT home page, MIT News, Imaging and Machine Vision Europe] [Industry Integration: NVIDIA, HuggingFace] [Code: GitHub (2.3k stars)]


Once for All: Train One Network and Specialize it for Efficient Deployment
OFA is an efficient AutoML technique that decouples model training from architecture search. Train only once, specialize for many hardware platforms, from CPU/GPU to hardware accelerators. OFA consistently outperforms SOTA NAS methods while reducing orders of magnitude GPU hours and CO2 emission. In particular, OFA achieves a new SOTA 80.0% ImageNet top1 accuracy under the mobile setting (<600M FLOPs). OFA is the winning solution for CVPR 2020 Workshop of Low-Power Computer Vision Challenge (FPGA track), 2019 IEEE Low-Power Image Recognition Challenge (classification and detection track), Low-Power Computer Vision Workshop at ICCV 2019 (DSP track). [Media: MIT News, Qualcomm, VentureBeat, SingularityHub] [Industry Integration: Meta, Sony, AMD] [Code: GitHub (1.9k stars), Colab]


ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware
ProxylessNAS is an efficient hardware-aware neural architecture search method, which can directly search on large-scale datasets. It can design specialized neural network architecture for different hardware platforms. With >74.5% top-1 accuracy, the latency of ProxylessNAS is 1.8x faster than MobileNetV2. [Media: MIT News, IEEE Spectrum] [Industry Integration: Meta, Amazon, Microsoft] [Code: GitHub (1.4k stars)]

Awards

Competition Awards

Academic Services

  • Serve as a reviewer for ICML, NeurIPS, ICLR, CVPR, ICCV, ECCV, TPAMI, IJCV, ACL, EMNLP