LeapLab@THU

discuss: huggingface.co/papers/2502.18…

Microsoft releases ART Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation

🔬 Read the paper: arxiv.org/abs/2409.00342 💻 Code & pretrained models: github.com/LeapLabTHU/Ada…

🚀 Excited to share our work on #ECCV2024: "AdaNAT: Exploring Adaptive Policy for Token-Based Image Generation". 🖼️ We introduce AdaNAT, a novel approach for efficient and high-quality image generation using adaptive policies in Non-autoregressive Transformers.

🔑 Key features: Learnable policy network for adaptive modulation of token generation Adversarial reward model for improved quality and diversity Significantly reduced inference time compared to diffusion models 📊 Impressive results on ImageNet, MSCOCO, and CC3M datasets!

very fantastic work done with @ALucyBrilliant @RayLu_THU @AndrewZ45732491. direct parameter editing can modulate LLM’s behavior

ConvLLaVA Hierarchical Backbones as Visual Encoder for Large Multimodal Models High-resolution Large Multimodal Models (LMMs) encounter the challenges of excessive visual tokens and quadratic visual complexity. Current high-resolution LMMs address the quadratic

📢Excited to share our recent work on Large Multimodal Models: ConvLLaVA. Without the encoding multiple image patches and multiple encoders, we use a hierarchical backbone, ConvNeXt, realizing high resolution understanding. arxiv.org/pdf/2405.15738

EfficientTrain++ is accepted by TPAMI2024🤩 🔥An off-the-shelf, easy-to-implement algorithm for training foundation visual backbones efficiently! 🔥1.5−3.0× lossless training/pre-training speedup on ImageNet-1K/22K! Paper&Code: arxiv.org/abs/2405.08768 github.com/LeapLabTHU/Eff…

Our recent work: Agent Attention! [High Performance & Linear Complexity] [Double the speed of SD and enhance generation quality, no additional fine-tuning is required] Paper and code: huggingface.co/papers/2312.08… arxiv.org/abs/2312.08874 github.com/LeapLabTHU/Age…

Excited to share our #NeurIPS2023 spotlight paper! 🌟 It proposes a novel offline-to-online RL algorithm, efficiently utilizing collected samples by training a family of policies offline and selecting suitable ones online. Check out our paper for details! arxiv.org/abs/2310.17966

ExpeL is now accepted at #AAAI24 ! The code and camera ready version will be updated promptly. Thanks for all the collaborators and see you in Vancouver!

Check us out at #NeurIPS2023 poster！We investigate into Q-value divergence phenomenon in offline RL and find self-excitation to be the main reason. Using layernorm in RL models can fundamentally prevent this from happening. arxiv.org/pdf/2310.04411…

Our recent work: Agent Attention! [High Performance & Linear Complexity] [Accelerate and improve Stable Diffusion, no additional fine-tuning is required] The paper and code have been released: arxiv.org/abs/2312.08874 github.com/LeapLabTHU/Age…