Feature 3DGS (Packaged Python Version)

This repo is the refactored Python training and inference code for Feature 3DGS. Built on top of gaussian-splatting, we reorganised the original code as a standard Python package with a modular Extractor-Decoder architecture, making it easy to swap foundation models without changing the core pipeline.

Each Gaussian point carries a learnable encoded semantics embedding alongside standard 3DGS attributes. A frozen Extractor produces ground-truth feature maps from training images, while a lightweight learnable Decoder maps the rasterised per-point embeddings back to the extractor's feature space. The decoder's per-point transform can also be applied directly to the stored embeddings, yielding extractor-aligned semantic features without rendering. The framework is backbone-agnostic: new foundation models can be plugged in by implementing an Extractor-Decoder pair and registering it.

Features

• Organised as a standard Python package with pip install support
• Modular Extractor-Decoder architecture for plugging in arbitrary foundation models
• Built-in DINOv3 support (ViT and ConvNeXt backbones)
• Built-in VGGT and VGG-T³ multi-view feature extractors
• Auto-registration pattern — add new models with zero changes to core code
• PCA-based feature visualisation for both ground-truth and rendered feature maps
• All training modes from upstream: base, densify, camera, camera-densify

Install

Prerequisites

• Pytorch (>= v2.4 recommended)
• CUDA Toolkit (12.4 recommended, match with PyTorch version)

Install dinov3 and vggt:

pip install --upgrade git+https://github.com/facebookresearch/dinov3.git@main
pip install --upgrade git+https://github.com/facebookresearch/vggt.git@main
pip install --upgrade Pillow hydra-core omegaconf # deps for vggt
pip install --upgrade git+https://github.com/jytime/LightGlue.git#egg=lightglue # deps for vggt

VGG-T³ support is bundled as a git submodule and currently requires a source checkout. Clone with --recursive as shown below, then install its runtime dependencies:

pip install -r submodules/vgg-ttt/requirements.txt

(Optional) If you have trouble with gaussian-splatting, try to install it from source:

pip install wheel setuptools
pip install --upgrade git+https://github.com/yindaheng98/gaussian-splatting.git@master --no-build-isolation

PyPI Install

pip install --upgrade feature-3dgs

or build latest from source:

pip install wheel setuptools
pip install --upgrade git+https://github.com/yindaheng98/feature-3dgs.git@main --no-build-isolation

Development Install

git clone --recursive https://github.com/yindaheng98/feature-3dgs.git
cd feature-3dgs
pip install --target . --upgrade . --no-deps

Download Checkpoints

DINOv3

Request access and download DINOv3 weights to checkpoints/:

checkpoints/
 ├── dinov3_convnext_base_pretrain_lvd1689m-801f2ba9.pth
 ├── dinov3_convnext_large_pretrain_lvd1689m-61fa432d.pth
 ├── dinov3_convnext_small_pretrain_lvd1689m-296db49d.pth
 ├── dinov3_convnext_tiny_pretrain_lvd1689m-21b726bb.pth
 ├── dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth
 ├── dinov3_vitb16_pretrain_lvd1689m-73cec8be.pth
 ├── dinov3_vith16plus_pretrain_lvd1689m-7c1da9a5.pth
 ├── dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth
 ├── dinov3_vits16_pretrain_lvd1689m-08c60483.pth
 ├── dinov3_vits16plus_pretrain_lvd1689m-4057cbaa.pth
 └── ...

VGGT

wget -P checkpoints/ https://huggingface.co/facebook/VGGT-1B-Commercial/resolve/main/vggt_1B_commercial.pt --header="Authorization: Bearer $HF_TOKEN"

VGG-T³

The vggttt extractor downloads nvidia/vgg-ttt through from_pretrained on first use. The VGG-T³ code and model are released under the NVIDIA OneWay Noncommercial License; consult the bundled submodule license before use.

Command-Line Usage

Visualise Extractor Output

Verify that the extractor produces meaningful features before training:

python -m feature_3dgs.show \
    --name dinov3_vitl16 \
    -s data/truck -d output/truck-dinov3_vitl16 \
    -o checkpoint_dir="'checkpoints'"

VGGT-family extractors process all views jointly and therefore require dataset cache preloading (enabled by default). The VGG-T³ extractors are CUDA-only. Available registered names are:

• vggt: final aggregator patch tokens, 2048 channels
• vggtrack: VGGT TrackHead DPT features, 128 channels
• vggttt: VGG-T³ final aggregator patch tokens, 2048 channels
• vggttttrack: VGG-T³ aggregation with the VGGT TrackHead, 128 channels

For example:

python -m feature_3dgs.show \
    --name vggttt \
    -s data/truck -d output/truck-vggttt

Train

python -m feature_3dgs.train \
    --name dinov3_vitl16 --embed_dim 32 \
    -s data/truck -d output/truck-semantic -i 30000 \
    --mode densify

Render

python -m feature_3dgs.render \
    --name dinov3_vitl16 --embed_dim 32 \
    -s data/truck -d output/truck-semantic -i 30000

Rendered feature maps are PCA-projected to RGB and saved alongside ground-truth feature visualisations.

Interactive Viewer

python -m feature_3dgs.viewer \
    --name dinov3_vitl16 --embed_dim 32 \
    -s data/truck -d output/truck-semantic -i 30000 \
    --port 8080

Opens an interactive viewer (via nerfview) that renders PCA-colourised semantic feature maps in real time from free-viewpoint camera controls.

API Usage

Dataset & Decoder

from feature_3dgs.prepare import prepare_dataset_and_decoder

dataset, decoder = prepare_dataset_and_decoder(
    name="dinov3_vitl16",   # registered extractor-decoder name
    source="data/truck",
    embed_dim=32,
    device="cuda",
)
# dataset is a FeatureCameraDataset; each camera carries a 'feature_map' in custom_data
# decoder is the learnable AbstractTrainableDecoder

Gaussian Model

from feature_3dgs.prepare import prepare_gaussians

gaussians = prepare_gaussians(
    decoder=decoder, sh_degree=3,
    source="data/truck", dataset=dataset, device="cuda",
)

SemanticGaussianModel extends GaussianModel with _encoded_semantics (per-point learnable embeddings in a compact latent space) and a _decoder. During rendering, the rasteriser splatts the encoded semantics into a 2D feature map, and the decoder transforms it to match the extractor's output space. The output dict contains both feature_map (decoded, extractor-aligned) and feature_map_encoded (raw rasterised).

Training

from feature_3dgs.prepare import prepare_trainer

trainer = prepare_trainer(gaussians, dataset, mode="densify")
for camera in dataset:
    loss, out = trainer.step(camera)

Inference

import torch
with torch.no_grad():
    for camera in dataset:
        out = gaussians(camera)
        rgb = out["render"]                  # (3, H, W)
        feat = out["feature_map"]            # (D, H', W')  decoded, extractor-aligned
        feat_enc = out["feature_map_encoded"] # (embed_dim, H, W)  raw rasterised

    # Per-Gaussian semantic features (no rendering needed)
    semantics = gaussians.get_semantics      # (N, D)  via decoder.decode_features

    # Custom linear projection at full resolution (e.g. PCA visualisation)
    weight, bias = ...  # (C, D) and (C,)
    out = gaussians.forward_projection(camera, weight, bias)
    projected = out["feature_map"]           # (C, H, W)

Save & Load

gaussians.save_ply("output/point_cloud.ply")
# also saves point_cloud.ply.semantic.pt and point_cloud.ply.decoder.pt

gaussians.load_ply("output/point_cloud.ply")

Design: Extractor & Decoder

The core abstraction decouples what features to distill (Extractor) from how to map rasterised embeddings back (Decoder).

Extractor (AbstractFeatureExtractor)

The extractor is a frozen foundation model that converts training images into dense feature maps. It runs only on the dataset side — each training view is processed once, cached, and served as the ground-truth supervision signal.

Image (C, H, W)  ──► Extractor (frozen) ──► Feature Map (D, H', W')

The extractor defines the target feature space (dimension D and spatial resolution H'×W'). It is never updated during training.

Decoder (AbstractSemanticDecoder)

The decoder is a learnable module with three core operations (defined on AbstractSemanticDecoder):

The trainable subclass AbstractTrainableDecoder adds:

Encoded semantics ──► Rasteriser ──► Raw Feature Map (embed_dim, H, W)
                                           │
                          ┌────────────────┼────────────────┐
                          ▼                ▼                ▼
                  decode_feature_map  forward_projection (stored as
                          │           (custom linear)    feature_map_encoded)
                          ▼                ▼
              Decoded Feature Map    Projected Map
               (D, H', W')          (C, H, W)

The default decode_feature_map applies decode_features per pixel (no spatial change). Subclasses may override it with reparameterized implementations for memory efficiency — e.g. the DINOv3 decoder reparameterizes a linear mapping followed by patch-level average pooling into a single F.conv2d call, avoiding a large intermediate tensor. Similarly, decode_feature_pixels reparameterizes two sequential linear layers into one combined projection.

The training loss is L1(Decoded Feature Map, Extractor Feature Map). The decoder's role is to bridge the gap between the compact per-point embedding (embed_dim, typically 32) and the extractor's high-dimensional output (D, e.g. 1024 for ViT-L), while also handling any spatial resolution change.

Why this split?

1. Memory efficiency: Only embed_dim channels are stored per Gaussian and rasterised, not the full D channels. The decoder upprojects after rasterisation.
2. Spatial alignment: Foundation models often output at patch resolution (e.g. 1/16 for ViT). The decoder can downsample the rasterised full-resolution map to match, avoiding expensive full-resolution feature supervision.
3. Direct feature access: decode_features can be applied directly to per-Gaussian encoded semantics (via get_semantics), producing extractor-aligned features without rendering.
4. Modularity: Swapping the foundation model only requires a new Extractor-Decoder pair. The Gaussian model, trainer, and rendering pipeline remain unchanged.

Extending: Adding a New Foundation Model

The project uses an auto-registration pattern. To add support for a new model (e.g. a hypothetical MyModel), follow the DINOv3 implementation as a reference:

Step 1: Implement the Extractor

Create feature_3dgs/mymodel/extractor.py:

import torch
from feature_3dgs.extractor import AbstractFeatureExtractor

class MyModelExtractor(AbstractFeatureExtractor):
    def __init__(self, model, ...):
        self.model = model
        self.model.eval()

    @torch.no_grad()
    def __call__(self, image: torch.Tensor) -> torch.Tensor:
        # image: (C, H, W) in [0, 1]
        # Return: (D, H', W') feature map
        ...

    def to(self, device) -> 'MyModelExtractor':
        self.model.to(device)
        return self

Step 2: Implement the Decoder

Create feature_3dgs/mymodel/decoder.py. The built-in LinearDecoder already provides a trainable nn.Linear, PCA-based initialisation, persistence, and reparameterized per-pixel operations. Subclass it and optionally override decode_feature_map / encode_feature_map if the extractor outputs at a different spatial resolution:

import torch
import torch.nn.functional as F
from feature_3dgs.decoder import LinearDecoder

class MyModelDecoder(LinearDecoder):
    def __init__(self, in_channels: int, out_channels: int, patch_size: int):
        super().__init__(in_channels, out_channels)
        self.patch_size = patch_size

    def decode_feature_map(self, feature_map: torch.Tensor) -> torch.Tensor:
        # Optional override for fused / memory-efficient implementation.
        # Default (inherited): applies decode_features per pixel (no spatial change).
        # Override to add spatial downsampling when the extractor outputs at
        # a lower resolution than the rasteriser (e.g. patch-level).
        ...

    def encode_feature_map(self, feature_map: torch.Tensor, camera) -> torch.Tensor:
        # Inverse of decode_feature_map. Override if decode_feature_map is overridden.
        ...

If your extractor outputs at the same resolution as the rasteriser, you can use LinearDecoder directly without subclassing.

The key design constraint: decode_feature_map's output spatial size and channel count must exactly match the extractor's output, so that L1 loss can be computed directly.

For example, the DINOv3 ViT extractor outputs at patch resolution (D, H/P, W/P). DINOv3LinearAvgDecoder reparameterizes a trainable nn.Linear with patch-level average pooling into a single F.conv2d call (kernel derived from linear weights, stride = patch size), avoiding the large (D, H, W) intermediate tensor entirely.

Step 3: Register via Factory

Create feature_3dgs/mymodel/registry.py:

from feature_3dgs.registry import register_extractor_decoder
from .extractor import MyModelExtractor
from .decoder import MyModelDecoder

FEATURE_DIM = 768  # D of your model's output

def factory(embed_dim: int, **configs):
    extractor = MyModelExtractor(...)
    decoder = MyModelDecoder(
        in_channels=embed_dim,
        out_channels=FEATURE_DIM,
        ...
    )
    return extractor, decoder

register_extractor_decoder("mymodel", factory)

Step 4: Trigger Registration on Import

Create feature_3dgs/mymodel/__init__.py:

from . import registry  # triggers register_extractor_decoder() at import time

Then add the import in feature_3dgs/__init__.py:

from . import mymodel  # auto-registers "mymodel"

After these steps, the new model is available everywhere:

python -m feature_3dgs.train --name mymodel --embed_dim 32 -s data/truck -d output/truck-mymodel -i 30000

Acknowledgement

This repo is developed based on Feature 3DGS, 3D Gaussian Splatting, and gaussian-splatting (packaged). Many thanks to the authors for open-sourcing their codebases.

Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields

Shijie Zhou, Haoran Chang*, Sicheng Jiang*, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, Achuta Kadambi (* indicates equal contribution)
| Webpage | Full Paper | Video | Original Code |

Abstract: 3D scene representations have gained immense popularity in recent years. Methods that use Neural Radiance fields are versatile for traditional tasks such as novel view synthesis. In recent times, some work has emerged that aims to extend the functionality of NeRF beyond view synthesis, for semantically aware tasks such as editing and segmentation using 3D feature field distillation from 2D foundation models. However, these methods have two major limitations: (a) they are limited by the rendering speed of NeRF pipelines, and (b) implicitly represented feature fields suffer from continuity artifacts reducing feature quality. Recently, 3D Gaussian Splatting has shown state-of-the-art performance on real-time radiance field rendering. In this work, we go one step further: in addition to radiance field rendering, we enable 3D Gaussian splatting on arbitrary-dimension semantic features via 2D foundation model distillation. This translation is not straightforward: naively incorporating feature fields in the 3DGS framework encounters significant challenges, notably the disparities in spatial resolution and channel consistency between RGB images and feature maps. We propose architectural and training changes to efficiently avert this problem. Our proposed method is general, and our experiments showcase novel view semantic segmentation, language-guided editing and segment anything through learning feature fields from state-of-the-art 2D foundation models such as SAM and CLIP-LSeg. Across experiments, our distillation method is able to provide comparable or better results, while being significantly faster to both train and render. Additionally, to the best of our knowledge, we are the first method to enable point and bounding-box prompting for radiance field manipulation, by leveraging the SAM model.

@inproceedings{zhou2024feature,
  title={Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields},
  author={Zhou, Shijie and Chang, Haoran and Jiang, Sicheng and Fan, Zhiwen and Zhu, Zehao and Xu, Dejia and Chari, Pradyumna and You, Suya and Wang, Zhangyang and Kadambi, Achuta},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21676--21685},
  year={2024}
}