Holi-Spatial¶

Development line: project:holi-spatial · thread holi-spatial-development
Last event: 2026-03-10 · 1 dated since 2026-03-10 · Researched: 2026-09-06 · confidence: medium

What it is¶

Holi-Spatial is a public data-preparation pipeline for research teams that need 3D Gaussian Splatting, masks, 3D boxes, and spatial QA from ScanNet, ScanNet++, or DL3DV scenes.

• Depth and point clouds: builds depth and point clouds with DA3, then runs PGSR/3DGS.
• Object annotation and QA: segments objects with SAM3 and a VLM, lifts masks to 3D, and generates QA.
• Multimodal dataset export: converts the output into an SFT dataset for LLaMA-Factory.

12K optimized 3DGS scenes and over 4M spatial annotations in the paper; the available HoliSpatial-2M-QA-Qwen3-VL-8B checkpoint is a BF16 model with 9B parameters and 17.6 GB.

This is an offline pipeline for curation and training with CUDA, external models, and your own scenes, not an interactive 3D service.

Development line¶

• 2026-03-10 — Holi-Spatial project resources were shared. On 2026-03-10, Holi-Spatial was linked to a public project page, source repository, and Hugging Face resource. This is a material public-development reference point, although the available evidence does not establish the precise release status or capabilities of the linked resource.

What changed¶

• 2026-03-08 — arXiv v1 published: describes a three-stage automated pipeline and Holi-Spatial-4M.
• 2026-03-10 — Holi-Spatial received a public project page and repository for video spatial annotation, claiming 13K+ scenes, 1.3M+ QA, and 300K+ 3D boxes.
• 2026-03-19 — the official Hugging Face page marks the ScanNetPP dataset as updated.
• 2026-03-21 — the official Hugging Face page marks the ScanNet_v2 dataset as updated.
• 2026-03-22 — the official Hugging Face page marks HoliSpatial-QA-2M as updated.
• 2026-03-29 — the official Hugging Face page marks HoliSpatial-2M-QA-Qwen3-VL-8B as an updated 9B checkpoint.
• 2026-04-21 — the official Hugging Face page marks HoliSpatial-3D-Grounding as updated.

How to use this¶

From 2026-03-10, practitioners should use the linked project page, repository, and Hugging Face resource as the starting point for evaluating Holi-Spatial, while independently verifying the exact artifact, version, and usage instructions.

1. Create a CUDA-enabled Python environment and install requirements with the two PGSR submodules. — https://github.com/Visionary-Laboratory/Holi-Spatial
2. Prepare one scene in the expected ScanNet, ScanNet++, or DL3DV structure; add covisibility root for QA. — https://github.com/Visionary-Laboratory/Holi-Spatial
3. Run DA3 on one scene through SCENE_DIR first; check depth_da3 and pointcloud_da3.ply. — https://github.com/Visionary-Laboratory/Holi-Spatial
4. Train PGSR/3DGS, then generate mesh and mesh-guided masks. — https://github.com/Visionary-Laboratory/Holi-Spatial
5. Start an OpenAI-compatible vLLM endpoint, run object/region annotation with SAM3, postprocess AABB, generate QA, and convert to LLaMA-Factory dataset.jsonl. — https://github.com/Visionary-Laboratory/Holi-Spatial

Best practices¶

• Start with one scene and check intermediate artifacts: batch scripts treat a scene as complete if a final JSON or trained point cloud already exists. — https://github.com/Visionary-Laboratory/Holi-Spatial
• Set up CUDA, Depth Anything 3, SAM3 assets, and an OpenAI-compatible vLLM endpoint explicitly instead of relying on implicit dependencies. — https://github.com/Visionary-Laboratory/Holi-Spatial
• Keep VLM verification for borderline instances: confidence filtering increases precision but can discard valid complex objects. — https://arxiv.org/abs/2603.07660
• Do not transfer results unverified to video with motion blur, heavy occlusion, few viewpoints, or dynamic objects: the authors list these as pipeline limits. — https://arxiv.org/abs/2603.07660

Superseded by this¶

• 2026-03-10 — The 'Data (coming soon)' and 'Benchmark (coming soon)' status on the project page no longer reflects availability: the repository reports a released data subset and checkpoint, and the Hugging Face organization now hosts four datasets and one model. This does not prove availability of the full Holi-Spatial-4M.

Still unknown¶

• https://hf.ru/linkd4c82 failed to open during verification; its destination URL, date, and contents are unconfirmed and unused.
• The March/April dates on the Hugging Face organization page are last-update dates rather than confirmed initial release dates: commit history pages were unavailable.
• The project page still lists data and benchmark as 'coming soon', while the README and Hugging Face show available artifacts. Whether this note refers to the full Holi-Spatial-4M or is simply outdated remains unclear.
• The checkpoint lacks a model card and a hosted inference provider; the exact prompt template, supported inputs, and recommended runtime remain unconfirmed.
• Neither the repository nor the model page lists an explicit license for code, datasets, or weights.

Sources¶

Agent brief¶

• Subject: project:holi-spatial, thread holi-spatial-development, 1 dated events 2026-03-10 → 2026-03-10.
• Practical note: From 2026-03-10, practitioners should use the linked project page, repository, and Hugging Face resource as the starting point for evaluating Holi-Spatial, while independently verifying the exact artifact, version, and usage instructions.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.