Trust3R — evidential uncertainty for feed-forward 3D reconstruction

Checkpoints for “Trust It or Not: Evidential Uncertainty for Feed-Forward 3D Reconstruction with Trust3R” (ICML 2026).

• Code: https://github.com/phai-lab/Trust3R
• Project page: https://trust3r-z.github.io/
• Paper: https://arxiv.org/abs/2605.19539

Trust3R adds two lightweight heads to a frozen MASt3R backbone: an evidential uncertainty head that predicts the parameters of a Normal-Inverse-Wishart prior over each 3D point — yielding a closed-form Student-t predictive distribution and a calibrated per-pixel uncertainty map in a single forward pass, no ensembles and no Monte Carlo sampling — and a gated residual head that applies small, gated corrections to the pretrained pointmap.

Files

File	Head	Backbone	Res.	Size
trust3r_niw_mast3r_224.pth	NIW evidential (full 3×3 covariance) + gated residual	frozen MASt3R ViT-L	224	3.0 GB
trust3r_nig_mast3r_224.pth	NIG evidential (diagonal variance) + gated residual	frozen MASt3R ViT-L	224	3.0 GB

NIW is the main model. NIG is the evidential-family ablation. Each .pth ships a .sha256 sidecar and a .metadata.json recording provenance, training mix and the evaluation protocol.

Download

pip install -U "huggingface_hub[cli]"
mkdir -p checkpoints
hf download phai-lab/Trust3R \
    trust3r_niw_mast3r_224.pth trust3r_nig_mast3r_224.pth \
    trust3r_niw_mast3r_224.pth.sha256 trust3r_nig_mast3r_224.pth.sha256 \
    --local-dir checkpoints/
(cd checkpoints && sha256sum -c *.sha256)

Usage

from mast3r.model import AsymmetricMASt3R

model = AsymmetricMASt3R.from_pretrained("checkpoints/trust3r_niw_mast3r_224.pth").eval()

The model expression is stored inside the checkpoint, so no architecture arguments are needed. A minimal pair-inference example is infer.py in the GitHub repo; the NIW predictive variance is recovered from the head outputs as

kappa = pred1["xyz_niw_kappa"]           # (1, 1, H, W)
nu    = pred1["xyz_niw_nu"]              # (1, 1, H, W)
Psi   = pred1["xyz_niw_Psi"]             # (1, 3, 3, H, W)
trace_Psi = Psi[:, 0, 0] + Psi[:, 1, 1] + Psi[:, 2, 2]
total_var = trace_Psi / (kappa.squeeze(1) * (nu.squeeze(1) - 4.0).clamp_min(1e-3))

Provenance and protocol metadata:

import torch
print(torch.load("checkpoints/trust3r_niw_mast3r_224.pth", map_location="cpu")["trust3r"])

Training

Initialised from MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth, backbone frozen, 150k steps at 224px (batch 10, 10 epochs) on a four-dataset mix of 150k pairs per epoch: ScanNet++ (25k), ARKitScenes (25k), Waymo (50k) and MegaDepth (50k). AdamW, base LR 3e-4 with cosine schedule, evidence regularisation λ_evi = 1e-3.

Evaluation

eval/reproduce_table1_table2.sh in the GitHub repo reproduces the paper tables from these checkpoints — AURC, AUSE, Spearman ρ, Sim(3)-aligned MAE/RMSE and NLL over ScanNet++, TUM RGB-D, KITTI and ETH3D. See eval/README.md there for the protocol.

License and intended use

CC BY-NC-SA 4.0 — non-commercial use only, inherited from MASt3R and DUSt3R. See CHECKPOINTS_NOTICE in the GitHub repo for the terms attached to the training datasets; ScanNet++, Waymo and ETH3D additionally require registration with their providers.

These weights are trained at 224px for research on uncertainty-aware 3D reconstruction. Other resolutions are outside the trained regime.

Citation

@misc{zhu2026trust3r,
  title  = {Trust It or Not: Evidential Uncertainty for Feed-Forward 3D Reconstruction with Trust3R},
  author = {Zhu, Zihao and Zhao, Wenyuan and Chen, Nuo and Tian, Chao and Fan, Zhiwen},
  year   = {2026},
  eprint = {2605.19539},
  archivePrefix = {arXiv}
}