nislamsm/CasualBench

Qwen-3D Cached ScanNet ViT Features

This dataset contains precomputed Qwen2.5-VL ViT features for ScanNet RGB frames, used by Qwen-3D as presented in Qwen-3D: A Generalist 3D Vision-Language Model for Spatial Understanding to skip the ViT forward pass during training.

Project page: https://qwen-3d.github.io/
Code: https://github.com/ll220/qwen3d
Model: https://huggingface.co/katefgroup/Qwen-3D
Paper: https://arxiv.org/abs/2608.02980

See docs/RUN.md in the code repository for full training instructions (CACHE_QWEN_FEATURES / FEATURE_DIR).

Contents

Archive	Backbone	Approx. size
scannet_image_qwen_features_3b.tar	Qwen2.5-VL-3B-Instruct	~182 GiB
scannet_image_qwen_features_7b.tar	Qwen2.5-VL-7B-Instruct	~319 GiB

Each archive includes a backbone prefix (3b/ or 7b/). After extracting both into the same parent directory, features are stored as bfloat16 tensors at:

FEATURE_DIR/
  3b/<scene_id>/<frame>.pt
  7b/<scene_id>/<frame>.pt

Example: FEATURE_DIR/3b/scene0051_03/140.pt.

Download and extract

# Download (or pass a single filename to download one archive)
hf download katefgroup/qwen3d_cached_scannet_vit_features --local-dir ./qwen3d_cached_scannet_vit_features

# Extract both archives into the same FEATURE_DIR
FEATURE_DIR="/path/to/scannet_image_qwen_features"
mkdir -p "$FEATURE_DIR"
tar -xf scannet_image_qwen_features_3b.tar -C "$FEATURE_DIR"
tar -xf scannet_image_qwen_features_7b.tar -C "$FEATURE_DIR"

This yields FEATURE_DIR/3b/... and FEATURE_DIR/7b/... (matching the layout used by Qwen-3D).

Usage with Qwen-3D

Point FEATURE_DIR at the parent directory (the code resolves 3b / 7b from QWEN_MODEL):

CACHE_QWEN_FEATURES True \
FEATURE_DIR "/path/to/scannet_image_qwen_features" \
QWEN_MODEL "Qwen/Qwen2.5-VL-3B-Instruct"

For the 7B backbone, keep the same FEATURE_DIR and switch the model:

CACHE_QWEN_FEATURES True \
FEATURE_DIR "/path/to/scannet_image_qwen_features" \
QWEN_MODEL "Qwen/Qwen2.5-VL-7B-Instruct"

• On a cache hit, the model skips the ViT forward.
• On a miss during ScanNet training, it runs the ViT and writes missing .pt files under the resolved backbone subdirectory.
• Caching applies to ScanNet datasets only. Leave CACHE_QWEN_FEATURES False (default) if you are not using cached features.

To rebuild the cache yourself instead of downloading, see docs/cache_qwen_vit_features.py.

Citation

@inproceedings{lin2026qwen3d,
  title     = {Qwen-3D: A Generalist 3D Vision-Language Model for Spatial Understanding},
  author    = {Lin, Lucy and Jain, Ayush and Liu, Yifan and Fragkiadaki, Katerina},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

License Note: This dataset is released under CC BY-NC 4.0. It may be used for non-commercial research purposes. Commercial use is not permitted under this license. ScanNet data itself is subject to the ScanNet Terms of Use.