Model Card for LHM v0.1

Overview

• This model card is for the LHM project, which is an official implementation of the paper LHM.
• Information contained in this model card corresponds to Version 0.1.

Model Details

• Training data

  Model	Training Data	Training Strategy
  LHM-MINI	2K2K & RP & THuman + 300K Videos	Random Crop Body Size
  LHM-500M	2K2K & RP & THuman + 300K Videos	Full Body
  LHM-500M-HF	2K2K & RP & THuman + 300K Videos	Random Crop Body Size
  LHM-1B	2K2K & RP & THuman + 300K Videos	Full Body
  LHM-1B-HF	2K2K & RP & THuman + 300K Videos	Random Crop Body Size

• Model architecture (version==0.3)

  Type	Layers	Feat. Dim	Attn. Heads	The number of GS Points.	Input Res.	Image Encoder	Encoder Dim.	Service Requirement
  LHM-MINI	2	1024	16	20K	512	dinov2_vits14_reg & Sapiens-1B	1024	16G GPU, 24G VRAM
  LHM-500M	5	1024	16	40K	512	dinov2_vits14_reg & Sapiens-1B	1024	18G GPU, 24G VRAM
  LHM-500M-HF	5	1024	16	40K	512	dinov2_vitb14_reg & Sapiens-1B	1024	18G GPU, 24G VRAM
  LHM-1B	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	22G GPU, 24G VRAM
  LHM-1B-HF	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	222G GPU, 24G VRAM

• Model architecture (with motion & save_memory; maximum supported length for 720P video is 20s)

  Type	Layers	Feat. Dim	Attn. Heads	The number of GS Points.	Input Res.	Image Encoder	Encoder Dim.	Service Requirement
  LHM-MINI	2	1024	16	20K	512	dinov2_vits14_reg & Sapiens-1B	1024	14G GPU, 24G VRAM
  LHM-500M	5	1024	16	40K	512	dinov2_vits14_reg & Sapiens-1B	1024	16G GPU, 24G VRAM
  LHM-500M-HF	5	1024	16	40K	512	dinov2_vitb14_reg & Sapiens-1B	1024	16 GPU, 24G VRAM
  LHM-1B	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	18G GPU, 24G VRAM
  LHM-1B-HF	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	18G GPU, 24G VRAM

• Model architecture (with motion)

  Type	Layers	Feat. Dim	Attn. Heads	The number of GS Points.	Input Res.	Image Encoder	Encoder Dim.	Service Requirement
  LHM-MINI	2	1024	16	20K	512	dinov2_vits14_reg & Sapiens-1B	1024	20G GPU, 24G VRAM
  LHM-500M	5	1024	16	40K	512	dinov2_vits14_reg & Sapiens-1B	1024	22G GPU, 24G VRAM
  LHM-500M-HF	5	1024	16	40K	512	dinov2_vitb14_reg & Sapiens-1B	1024	22 GPU, 24G VRAM
  LHM-1B	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	28G GPU, 24G VRAM
  LHM-1B-HF	15	1024	16	40K	1024	dinov2_vitb14_reg & Sapiens-1B	1024	28G GPU, 24G VRAM

License

• The model weights are released under the Creative Commons Attribution-NonCommercial 4.0 International License.
• They are provided for research purposes only, and CANNOT be used commercially.