SplatAnything wins this comparison. In the authors' own seven-dataset benchmark it ranks first on most datasets and settings against Mirror2-GS and other feed-forward rivals, and its weights and inference code went public on 8 November 2026 (paper, GitHub). It is the better pick for prototypes and products built outside the EU, UK and South Korea, provided your platform stays under 1 million monthly active users. Anyone else needs to read the licence before writing a line of integration code.
This is a practical look at SplatAnything 3D gaussian splatting open weights: what changed technically, how it compares, what it needs to run, and the licence detail that the Hugging Face model card does not make obvious.
TL;DR
- SplatAnything is a feed-forward 3D Gaussian Splatting model from Zhejiang University and Tencent Hunyuan. It builds a splat scene in seconds, with no per-scene optimisation (paper).
- Against Mirror2-GS, the next-best method in the paper's tables, it scores clearly better on the DTU dataset (details below).
- The weights total 16.56 GB on Hugging Face. No official VRAM figure has been published.
- The GitHub licence excludes the EU, UK and South Korea, and requires a separate Tencent licence above 1 million monthly active users.
- The Hugging Face card is tagged Apache-2.0, which conflicts with the GitHub licence file. Treat the stricter terms as the ones that apply.
- A free hosted demo exists if you want to test before downloading anything.
What problem does SplatAnything fix in feed-forward splatting?
Earlier feed-forward 3DGS models predict Gaussian attributes in a space normalised by radial distance, called RDNorm. The paper shows this leaves a gap: the spacing between neighbouring Gaussian centres still varies by up to 4.5x across scene categories (paper, Fig. 4 and Sec. 3.3). The network therefore receives ambiguous supervision. In practice that shows up as holes in small-object scenes and over-smoothing in large outdoor ones.
The proposed fix is Gaussian Spacing Normalization (GSNorm). Per-pixel spacing is roughly depth divided by focal length. GSNorm takes a robust scene-level statistic of that spacing, the median, and canonicalises it before attributes are predicted. The architecture pairs a frozen WorldMirror 2.0 geometry backbone with a lightweight GSHead, and Gaussian centres are fixed to back-projected geometry.
How does SplatAnything compare with Mirror2-GS on benchmarks?
The authors built a cross-domain benchmark covering seven datasets: ScanNet, NRGBD, DL3DV, Tanks & Temples, Mip-NeRF 360, ETH3D and DTU. It tests 1, 4, 8, 16, 32 and 64 input views, with input-target overlap from 0.2 to 0.6 and resolutions from roughly 480p to 1K. Metrics are LPIPS and DISTS, where lower is better. SplatAnything ranks first on most datasets and settings, both without and with camera poses (paper, Tables 1 to 3).
Two readings from the tables, using the DTU dataset averaged over 4 to 64 views:
| Setting | Metric | SplatAnything | Mirror2-GS |
|---|---|---|---|
| Without pose | DISTS | 0.229 | 0.347 |
| Without pose | LPIPS | 0.483 | 0.617 |
| With pose | DISTS | 0.184 | 0.315 |
| With pose | LPIPS | 0.364 | 0.532 |
The ablation is the more useful result for understanding why. Averaged across datasets, GSNorm beat RDNorm at every view count. At one view, DISTS was 0.294 against 0.308 and LPIPS was 0.367 against 0.394. At 64 views, DISTS was 0.217 against 0.224 and LPIPS was 0.229 against 0.250 (paper, Table 3). The gap is modest at DISTS and larger at LPIPS, so expect a consistent improvement rather than a dramatic one on every scene.
One caveat: these are the authors' benchmark and the authors' choice of baselines. Independent reproduction has only just become possible now that the weights are public.
What hardware does SplatAnything need to run?
The official README asks for Python 3.10 or later and PyTorch 2.4 or later with a CUDA build, so an NVIDIA GPU is expected (GitHub). It does not publish a VRAM requirement.
The download is 16.56 GB in total: a 11.33 GB SplatAnything checkpoint, a 5.05 GB WorldMirror 2.0 file and a 0.18 GB sky-segmentation model (Hugging Face). From that size alone, our inference is that a 12 GB card is unlikely to cope, a 16 GB card such as a free Colab T4 is marginal, and 24 GB should be comfortable. This is an estimate, not a measurement.
For a related data point, a community ComfyUI wrapper for the earlier InfiniSplat line recorded 1.03 GiB peak CUDA allocation on a 24 GB laptop RTX 5090 for single-image inference (ComfyUI-InfiniSplat). That is a different model revision, so it should not be read as a SplatAnything number.
If you would rather not guess, the hosted demo Space runs in the browser.
How do you run SplatAnything on your own images?
Per the README, put overlapping RGB images in an images/ folder inside a scene directory. A prior_camera.json is optional. Then run:
python inference.py --input_path examples/object --mode rgb --output_path outputs/object
--mode rgb predicts the cameras itself. --mode pose uses cameras you supply. The model input resolution is set with --resolution W H, defaults to 756×756, and must be a positive multiple of 14, with 1K and above supported (inference README). --sky_seg enables optional sky removal through the ONNX model, and --opacity_thresh filters low-opacity Gaussians from the exported PLY file, which any standard splat viewer can open. The model accepts single images, multi-camera rigs and panoramas, and the paper positions it for driving, robot navigation, AR/VR and game development.
Is SplatAnything free for commercial use?
Partly. The licence file in the GitHub repository is the Tencent HunyuanWorld-Mirror Community License Agreement, which GitHub's API reports as unrecognised rather than a standard open licence. Two clauses matter (GitHub):
- The licensed territory is worldwide excluding the EU, UK and South Korea.
- If your products or services had more than 1 million monthly active users in the calendar month before the licence's own release date of 22 October 2025, you must request a licence from Tencent, which it may grant or refuse at its sole discretion.
The Hugging Face model card is tagged apache-2.0. That tag conflicts with the repository, and the checkpoint bundles the WorldMirror 2.0 weights, which are the source of the stricter terms. Anyone who assumed Apache from the tag alone could ship into a restricted territory by mistake. This is not legal advice; have counsel check the actual text if you plan to ship.
Why is SplatAnything trending now?
It reached the top of the Hugging Face daily papers list on 9 November 2026 with 61 upvotes, according to a daily-paper tracker. It also lands in a busy stretch for open releases: one count found 21 verified open-weight language model releases between 1 September and 7 November 2026 (GPUwerk). Licence fine print is a recurring theme across that wave. Our coverage of weight-space editing, the CUA-Master 27B computer-use agent, the GroundingPi 4B model, PLaMo 3 Translate and Arete 27B shows how much terms vary between releases that all describe themselves as open.
FAQ
Q: Is SplatAnything free for commercial use?
A: Largely, outside the EU, UK and South Korea, and below 1 million monthly active users (GitHub licence). Above that threshold you must request a licence from Tencent, and the territory exclusion applies regardless of size.
Q: Does SplatAnything run on a 12 GB GPU?
A: The project has not published a VRAM figure. Given 16.56 GB of checkpoints, 12 GB looks unlikely, 16 GB is marginal and 24 GB should be comfortable. The hosted demo avoids the question.
Q: What does SplatAnything beat?
A: In the authors' seven-dataset benchmark it ranks first on most datasets and settings against recent feed-forward methods including Mirror2-GS, AnySplat and YoNoSplat, measured with LPIPS and DISTS.
Q: Does it work with a single image or only multiple views?
A: Both. The paper benchmarks 1, 4, 8, 16, 32 and 64 input views and includes a comparison on single-image input against SHARP and InfiniSplat.
Q: Why does the Hugging Face page say Apache-2.0?
A: The model card carries an Apache-2.0 tag, but the GitHub licence file is Tencent's community licence with territory and user-count limits. Follow the stricter file and confirm with counsel.
Last verified: 10 November 2026. Primary sources: arXiv abstract, full text, GitHub repository. Corrections log: none yet. This article was produced with AI assistance and reviewed against the cited sources.
Top comments (0)