topics of interest

The workshop targets computational efficiency of training, adaptation, and inference for large vision models including but not limited to:

• Scalable architectures

  • Long context modeling by developing state space models e.g., Mamba, and recurrent models e.g., RWKV, GLA and xLSTM, for computer vision

  • Scaling up model parameters by mixture of experts architectures e.g., sparse MoE for computer vision

  • Optimizing vision transformers e.g., token merging and pruning, and approximated self-attention

• Efficiency of generative models

  • Efficient architectures e.g., hierarchical diffusion transformers (DiTs), scalable pixel-space diffusion models, and efficient auto-regressive models

  • Fast diffusion sampling e.g., adversarial distillation, and consistency models

  • Efficient autoregressive models for generating images and videos, e.g., parallel decoding, speculative decoding, and KV caching for vision models

  • Efficient generation of high-resolution images and videos

  • Efficient generation of 3D objects and scenes e.g., by diffusion models, Gaussian splatting, and NeRF

  • Efficient generation of vision language models e.g. multi-modal LLMs

• Efficiency of foundation vision models

  • Efficient foundation models for perception e.g., open set object detection, semantic segmentation and depth estimation

  • Efficient foundation models for representation learning e.g., DINO and CLIP

• Compression, quantization, conditional compute, pruning and distillation of LVMs

• Deploying LVMs on low power devices e.g., smartphone