YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
TriSplat++ model artifacts
The reference LGTM appearance model is:
/root/data/haoxuan/TriSplat/outputs/exp_lgtm10k_train/2026-08-28_11-58-26/checkpoints/render_step_002700.ckpt
It is a 149 MiB Lightning checkpoint (global step 2700) containing only the
trained DA3/TSDPT output head. Its state_dict has 92 tensors under
encoder.da3.model.gs_head.*; it does not contain the DA3-GIANT backbone,
camera encoder, depth/feature branches, or a complete standalone model. The
LGTM texture-remapping implementation lives in the repository and must be
constructed by the TriSplat++ code path.
This is therefore a head-only artifact and must be paired with the complete
DA3-GIANT-1.1 checkpoint at runtime. Point DA3_CHECKPOINT (or the
renderer’s --da3-checkpoint) at the full DA3 model directory, for example:
export DA3_CHECKPOINT=/path/to/Depth-Anything-3/checkpoints/DA3-GIANT-1.1
export TRISPLATPP_CHECKPOINT=$PWD/weights/trisplatpp_lgtm_step2700.ckpt
The SHA-256 digest of this head-only artifact is:
9fa4c3011bf8aff893a18768330855a4c27b19c413c3ff11cba40ad653331981
The binary is tracked with Git LFS because GitHub's normal object limit is 100 MiB. If the remote does not provide LFS, keep the artifact outside Git and export its path instead:
export TRISPLATPP_CHECKPOINT=/path/to/render_step_002700.ckpt
The checked-in LFS filename is weights/trisplatpp_lgtm_step2700.ckpt. The
same file is published in the Hugging Face model repository
hhx112340/trisplatplus.