Quantitative comparison

on Open3DHOI^[4] in the wild

LEXIS-Flow sets SotA on in-the-wild Open3DHOI^[4].

Generative vs learning-based
init object shape
(SAM3D)

Refinement vs optimization-based
'Expert' init for all
(CameraHMR & SAM3D)

Method	$\mathrm{CD}_{\mathrm{hum}}$ ↓	$\mathrm{CD}_{\mathrm{obj}}$ ↓	Collision ↓	Contact ↑
AHDM^[7] (w. scale align.)	13.50	49.38	0.089	0.141
BLEXIS-Flow (Ours)	8.85	35.01	0.060	0.211

CInteractVLM^[6]	7.20	38.20	0.054	0.372
DCameraHMR^[1] + SAM3D^[2]	7.20	37.30	0.051	0.182
EHOI-Gaussian^[5]	7.28	32.02	0.061	0.151
FInteractVLM++^[6]	7.20	30.11	0.047	0.394
GLEXIS-Flow* (Ours)	7.05	22.96	0.041	0.451

Ours best on all metrics

[1] CameraHMR, Patel et al 3DV'25

[2] SAM3D, SAM3D team arXiv'25

[4] Open3DHOI, Wen et al CVPR'25

[5] HOI-Gaussian, Wen et al CVPR'25

[6] InteractVLM, Dwivedi et al CVPR'25

[7] HDM, Xie et al CVPR'24