T4GS — LSeg feature lifting
Inference weights for timestamp-conditioned 3D Gaussians with 512-dimensional LSeg features. The companion code renders full feature maps and exports PCA visualizations; no evaluation datasets or semantic annotations are needed.
Files
| File | Contents |
|---|---|
t4gs_lseg.ckpt |
Complete stage-2 encoder state_dict (keys prefixed encoder.) and global_step=5000; optimizer state omitted. |
demo_e200.ckpt |
LSeg state_dict, including ViT, decoder, and CLIP parameters. Legacy training metadata omitted. |
checksums.json |
SHA-256 and byte sizes for both files. |
The Gaussian feature dimension is 512, with 2048 Gaussians, 224×224 context
images, sinusoidal timestamp period 10000, and target_1 time normalization.
Use the matching config/inference/lseg.yaml supplied with the code. These
settings differ from the reconstruction-only T4GS release.
Download
From the companion code directory:
python scripts/download_lseg_weights.py
Or download directly:
from huggingface_hub import hf_hub_download
for name in ("t4gs_lseg.ckpt", "demo_e200.ckpt"):
hf_hub_download("honggyuAn/T4GS-LSeg", name, local_dir="pretrained_weights")
Run
Follow the companion code README for environment installation, then run:
python scripts/run_lseg.py --input /path/to/frames --output outputs/lseg --render-pca
Input: JPG/PNG frame folder, lexicographically ordered. Up to 12 context frames
are sampled and center-cropped to 224×224. Outputs: per-timestamp Gaussian
files, rendered 512-D feature maps, RGB images and PCA images. The default
renderer uses identity camera-to-world and approximate normalized focal 0.8767;
provide --cameras cameras.npz for your own normalized cameras. PCA colors are
feature visualizations, not semantic class predictions or evaluation scores.
The code is supplied separately as T4GS-LSeg-code.zip, with README.md and
README_LSEG.md. This model repository contains weights and metadata only.
Provenance and license
The lifting weights were exported without changing model tensors from the
exp_feature_rendering_lseg_5dataset_lpf experiment, run 2026-05-05, step 5000.
LSeg weights were exported from the existing demo_e200.ckpt distribution used
by the source implementation. Both exports load with torch.load(..., weights_only=True).
T4GS/C3G/VGGT-derived weights retain the project's CC BY-NC 4.0 terms.
LSeg implementation: https://github.com/isl-org/lang-seg (MIT).
OpenAI CLIP implementation: https://github.com/openai/CLIP (MIT).
Third-party components retain their original terms; this card does not
relicense their original weights. See the companion THIRD_PARTY_LICENSES.md.
These weights are for feature lifting inference. No new benchmark performance claims are made by this export.