AnySplat
AnySplat (InternRobotics): 3D Gaussian splats, cameras and depth from unposed images in one pass; saves a 3DGS .ply. MIT.
experimental Fork of InternRobotics/AnySplat@5f5e208 · backends: portable, kernels
- Paper AnySplat: Feed-forward 3D Gaussian Splatting from Unconstrained Views arXiv · first released 29 May 2025
Checkpoint anysplat
Paper 2025 Added in 0.2.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import multiview
model = multiview.load("anysplat") # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth # also: confidence, world_points, intrinsics, cam_to_world, gaussians
result.backend # what ran: mode, device, dtype, kernels fastforks bench anysplat Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
Input: synthetic, 336x518
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original (own default)the authors' mixed-precision option (bfloat16); InternRobotics/AnySplat@5f5e208a7dd5 EncoderAnySplat inference, long side 448, voxel_size 0.002 | bf16 | 741 | – | 8,931 MiB/9,394 MiB | 7.0 | 0.82 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 503 | 1.5× | 4,842 MiB/4,932 MiB | 4.7 | 0.89 |
| Kernels | fp16 | 364 | 2.0× | 4,849 MiB/5,192 MiB | 4.0 | 0.71 |
| Portable + torch.compile | 427 | 1.7× | 4,137 MiB/4,694 MiB | 80 (build) | – | |
| TensorRT fp16 | fp16 | 314 | 2.4× | 4,504 MiBenginenot comparable | 215 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-12 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | cannot run here | – | 336x518 | |
| Portable | fp16 | 5444 | 8,751 MiB process | matches | 336x518 | |
| Portable | fp32 | 33262 | 10.5 GiB process | matches | 336x518 | |
| Kernels | fp16 | not implemented | – | 336x518 | ||
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- MIT AND LicenseRef-VGGT LICENSELICENSE-VGGT
- Weights
- MIT source
- Training data
- Per the paper (arXiv 2505.23716) a mixture of public multi-view datasets; the repository ships CO3Dv2, DL3DV and ScanNet++ training configs. VGGT-1B initialisation.
- AnySplat's code and checkpoint are MIT; its VGGT copy (net/) is under the VGGT License (commercial use allowed). The checkpoint was fine-tuned from VGGT-1B, whose own weights are CC-BY-NC-4.0: the AnySplat authors release theirs as MIT.
- Upstream's Gaussian head file (vggt_dpt_gs_head.py) carries a stale Naver CC BY-NC-SA header from DUSt3R although the class used here is new code on the VGGT DPT head; this family reimplements that head (GSHead in model.py) and vendors none of the Naver-headed files.
Our reading of the licences, not legal advice.
Weights
model.safetensors 2.9 GB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download anysplat).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
DTU scans 1
9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 %
| Run | dtype | AbsRel ↓ | δ1.25 % ↑ | Δ AbsRel vs fp32 original | scans | ms / scan |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0108 | 99.93 | reference | 5 | 2352 |
| Original fp16 | bf16 | 0.0108 | 99.93 | ±0.0000 | 5 | 956 |
| Portable fp32 | fp32 | 0.0108 | 99.93 | ±0.0000 | 5 | 2345 |
| Portable | fp16 | 0.0109 | 99.93 | −0.0001 | 5 | 645 |
| Kernels | fp16 | 0.0108 | 99.93 | ±0.0000 | 5 | 461 |
640x512 crops, the 7 most co-visible views as sources; GT = depth rendered from the structured-light points. AbsRel / delta1 over GT pixels, after per-image median scaling unless the model takes the cameras (calibrated MVS: cameras, error in mm; PatchMatchNet also the depth range 425-935 mm, MVSAnywhere its own range from the camera baselines). Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12; portable fp32, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 62af007c, 2026-11-12.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 1.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0013 and p99 <= 0.0033 (absolute error of the output depth).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.000013 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.0013 | 0.00082 |
| GPU default dtypes, p99 |diff| | ≤ 0.0033 | 0.00214 |
Recorded for anysplat on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: cudnn_haccfastbb
Details
AnySplat (Jiang et al. 2025): feed-forward 3D Gaussian splatting from unposed, uncalibrated images. A VGGT backbone (fine-tuned aggregator, camera head and depth head) predicts the cameras and depth of every view; a DPT Gaussian head predicts per-pixel Gaussians (opacity, scale, rotation, degree-4 SH), fused per voxel into one scene (`result.gaussians`, `.save("scene.ply")`). Forward pass only: no renderer.
- Input / output
- Input: several images of one scene (posed and calibrated for multi-view stereo, unposed otherwise). Output: depth maps and, where the model predicts them, cameras and 3-D points.
- Input size
- any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (6,627 MiB allocated)
- Options
long_side=448>= 14· changes the outputvoxel_size=0.002>= 0.0· changes the output
- Outputs
depth,confidence,world_points,intrinsics,cam_to_world,gaussians- CUDA kernels
fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonescudnn_haccconvolutions with fp16 accumulation (cuDNN)
- Apple (MPS)
- fp32, fp16
Limitations
- Kernels peak allocated is equal to portable, but the cuDNN workspace next to the large conv output (257 MiB in the DPT head) fragments the allocator: peak reserved is ~260 MiB above portable.
Full reference, generated from the repository
AnySplat (InternRobotics): 3D Gaussian splats, cameras and depth from unposed images in one pass; saves a 3DGS .ply. MIT.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
anysplat | synthetic (336x518) | original portable kernels portable + torch.compile TensorRT fp16 | 741 503 364 427 314 | 8931 / 9394 4842 / 4932 4849 / 5192 4137 / 4694 7629 / 7674 |
Details
- Original code: https://github.com/InternRobotics/AnySplat at
5f5e208 - Code licence: MIT AND LicenseRef-VGGT (permissive); files:
LICENSE,LICENSE-VGGT - Weights licence: MIT
- Licence evidence: https://github.com/InternRobotics/AnySplat/blob/5f5e208a7dd57d52e43ea0d553a95eab526e8775/LICENSE and https://huggingface.co/lhjiang/anysplat/blob/d2e8c343672646041ad4ea518184968f94362f01/README.md (license: mit)
- Training data: Per the paper (arXiv 2505.23716) a mixture of public multi-view datasets; the repository ships CO3Dv2, DL3DV and ScanNet++ training configs. VGGT-1B initialisation.
- Licence note: AnySplat's code and checkpoint are MIT; its VGGT copy (net/) is under the VGGT License (commercial use allowed). The checkpoint was fine-tuned from VGGT-1B, whose own weights are CC-BY-NC-4.0: the AnySplat authors release theirs as MIT.
- Licence note: Upstream's Gaussian head file (vggt_dpt_gs_head.py) carries a stale Naver CC BY-NC-SA header from DUSt3R although the class used here is new code on the VGGT DPT head; this family reimplements that head (GSHead in model.py) and vendors none of the Naver-headed files.
- Options:
long_side= 448 (>= 14),voxel_size= 0.002 (>= 0.0) - Kernel libraries:
fastbb,cudnn_hacc - Variants:
anysplat - Outputs: depth, confidence, world_points, intrinsics, cam_to_world, gaussians; kind:
scale_invariant - Input size: any H x W >= 1, resized or padded inside the model
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
anysplat3000x4000 (6.5 GiB peak allocated, 9.4 reserved) anysplat336x518, first call: portable + torch.compile 80 s (cold Inductor cache), TensorRT fp16 215 s (ONNX + engine).anysplatoriginal 336x518: the authors' mixed-precision option (bfloat16); InternRobotics/AnySplat@5f5e208a7dd5 EncoderAnySplat inference, long side 448, voxel_size 0.002anysplatload / first call: original 8.32 s / 3.068 s (336x518), portable 4.71 s / 0.891 s (336x518), kernels 4.04 s / 0.708 s (336x518)- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
640x512 crops, the 7 most co-visible views as sources; GT = depth rendered from the structured-light points. AbsRel / delta1 over GT pixels, after per-image median scaling unless the model takes the cameras (calibrated MVS: cameras, error in mm; PatchMatchNet also the depth range 425-935 mm, MVSAnywhere its own range from the camera baselines). Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|---|
anysplat | DTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % | 0.0108 / 99.93 | 0.0108 / 99.93 | 0.0109 / 99.93 | 0.0108 / 99.93 | 0.0108 / 99.93 |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 62af007c, 2026-11-12.
Parity
CPU, float32: max |diff| <= 1.3e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0013 and p99 <= 0.0033 (absolute error of the output depth).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
anysplat | 0.00082 / 0.00214 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fastbb@7d058322634d:sm_120 |
Limitations
- Kernels peak allocated is equal to portable, but the cuDNN workspace next to the large conv output (257 MiB in the DPT head) fragments the allocator: peak reserved is ~260 MiB above portable.
Example output
No example rendered for this checkpoint yet.