leeiioolin/ViPS

Beyond Single Expert: Harmonizing Diverse Visual Priors in MLLMs for Spatial Understanding

NeurIPS 2026

Xiao Lin, Xiaohu Huang, Kai Han

Visual AI Lab, The University of Hong Kong

Paper · Project page · Code

Checkpoints

Location	Description
vips/	ViPS checkpoint.
vggt_multihead.pt	Prior proxy checkpoint for prior feature extraction.

Download

ViPS requires Qwen3-VL-8B-Instruct. Run these commands from the code repository root:

hf download leeiioolin/ViPS --local-dir checkpoints
hf download Qwen/Qwen3-VL-8B-Instruct --local-dir models/Qwen3-VL-8B-Instruct

See the code README for setup and usage.

Evaluation of the ViPS checkpoint

VSI-Bench results:

Model	Overall	Obj. Count	Abs. Dist.	Obj. Size	Room Size	Rel. Dist.	Rel. Dir.	Route Plan	Appr. Order
ViPS	63.7	69.6	54.7	71.1	72.5	60.1	81.9	45.9	53.6

License

Component	License
Original ViPS code and original adapter / fusion-module contributions	MIT
Qwen-derived components	Apache-2.0
Prior Proxy (vggt_multihead.pt, derived from VGGT-1B)	CC BY-NC 4.0 — non-commercial

See LICENSE for attribution, scope, and full terms.

Citation

@article{Lin2026ViPS,
  author  = {Xiao Lin and Xiaohu Huang and Kai Han},
  title   = {Beyond Single Expert: Harmonizing Diverse Visual Priors in MLLMs for Spatial Understanding},
  journal = {arXiv preprint arXiv:2607.15054},
  year    = {2026}
}