F⁴Splat¶

Development line: project:f4splat · thread f4splat-development
Events: 2 dated, 2026-03-30 → 2026-07-24 · Researched: 2026-09-04 · confidence: medium

What it is¶

F⁴Splat is a 3DGS research framework for developers synthesizing novel views from sparse uncalibrated images. It is a code alternative to AnySplat for CUDA research, not a finished cloud service or universal 3D scanner.

• Camera and parameter prediction: predicts camera poses, Gaussian parameters, and densification score maps.
• Primitive allocation: distributes Gaussian primitives by scene complexity and cross-view overlaps.
• Budget scaling: changes the target Gaussian budget at inference without retraining.

The local demo takes 2–16 images; published weights cover ACID 2-view and RealEstate10K 2-/24-view.

Development line¶

• 2026-03-30 — F⁴Splat project page published. F⁴Splat appeared on its public project page on 2026-03-30. A reference to level-of-detail graphics framed the announcement in rendering development. The page gave no exact technical claims or release status.
• 2026-07-24 — F⁴Splat GitHub repository linked. The development line linked the F⁴Splat GitHub repository on 2026-07-24 alongside the March announcement. This opened public access to source code. The post did not specify which code, version, or implementation changes appeared on that date.

What changed¶

• 2026-03-30: The project page described the method. Instead of uniform pixel- or voxel-to-Gaussian distribution, the model allocates primitives by predicted density.
• 2026-06-30: The official commit "Release public code" added 181 files with demo, training, evaluation, and source code. The method became a reproducible local workflow.
• 2026-07-24: The link pointed to the official repository, but the source recorded no separate technical change on that day. The nearest confirmed code release remains 30 June.
• 2026-09-03: The arXiv paper updated to v3. The page provides no description of differences, so we cannot claim new features or altered results from this revision.

How to use this¶

As of 2026-07-24, consult the F⁴Splat project page for context and the linked GitHub repository for source access. Treat precise capabilities and version status as unverified pending research.

1. Clone the repository, create a Python 3.11 environment, install PyTorch 2.8.0 for CUDA 12.8, and install project dependencies. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md
2. Download the Hub checkpoint matching the target dataset and context view count: acid-2view, re10k-2view, or re10k-24view. — https://huggingface.co/Knowing/F4Splat
3. Run python demo/app.py, upload 2–16 images, and select a Gaussian budget to compare renders. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md
4. For experiments, prepare the dataset in PyTorch chunk format with index.json, then run the supplied training or evaluation scripts with GPU ID and checkpoint path. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md

Best practices¶

• Match the PyTorch wheel to the installed CUDA version first. If CUDA is not 12.8, install matching PyTorch before other dependencies. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md
• Choose checkpoints by dataset and context view count. Do not treat published weights as a universal model for arbitrary scenes. — https://huggingface.co/Knowing/F4Splat
• Evaluate the Gaussian budget by actual primitive count. The demo renders interpolation-only video and marks results with true Gaussian counts in thousands. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md
• Follow the license and access terms for RealEstate10K, DL3DV, and ACID when training. — https://raw.githubusercontent.com/mlvlab/F4Splat/main/README.md

Superseded by this¶

• 2026-06-30: Treating F⁴Splat solely as a paper or project page is obsolete. Official source code is available with demo, training, and evaluation scripts.
• 2026-09-03: Cite arXiv v3 instead of v1/v2 for current research citations. This revision does not demonstrate runtime capability changes.

Still unknown¶

• Primary source text was not readable, leaving any separate release claim for 2026-07-24 unextracted.
• The 2026-09-03 arXiv v3 preprint includes no change note, leaving differences from v1/v2 unconfirmed.
• Official materials provide no independent quality evaluation on user captures, runtime speed, or GPU memory usage.

Sources¶

Agent brief¶

• Subject: project:f4splat, thread f4splat-development, 2 dated events 2026-03-30 → 2026-07-24.
• Practical note: As of 2026-07-24, consult the F⁴Splat project page for context and the linked GitHub repository for source access, while treating precise capabilities and version status as unverified pending research.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.