DispViT+ ViT-G
Full DispViT+ stereo-disparity model from DispViT: Direct Stereo Disparity Regression with a Single-Stream Vision Transformer (ICLR 2026).
The checkpoint contains the frozen DispViT ViT-G regressor, the ConvNeXt feature
extractor, and all learned refinement weights. Its embedded model_config
reconstructs the complete architecture through the current
DispViTPlus.from_pretrained API. Separate regressor or backbone downloads are
not required.
Architecture
| Setting | Value |
|---|---|
| Parameters | 1,282,170,176 |
| Regressor | DispViT ViT-G, 8 channel groups |
| Feature extractor | ConvNeXt, 256 output channels |
| Refinement | 6 blocks, 256-dimensional embeddings |
| Attention | 4 heads, window size 6 |
| Saved weights | Raw final model state |
Usage
Use a checkout of the DispViT repository
that provides dispvit.models.dispvit_plus.DispViTPlus.from_pretrained, and
install its documented inference dependencies. Run this example from that
checkout:
import numpy as np
from PIL import Image
import torch
from dispvit.models.dispvit_plus import DispViTPlus
device = "cuda" if torch.cuda.is_available() else "cpu"
model = DispViTPlus.from_pretrained("aeolusguan/dispvit-plus-vitg").to(device).eval()
def read_image(path):
image = np.array(Image.open(path).convert("RGB"), copy=True)
return torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float().to(device)
batch = {"img1": read_image("left.png"), "img2": read_image("right.png")}
with torch.inference_mode():
predictions = model(batch)
disparity = predictions["disp"] # [B, H, W], in pixels
Inputs are rectified left/right RGB images with matching dimensions, represented
as [B, 3, H, W] tensors with values in [0, 255]. The model pads images
internally and returns the final disparity at the original image size.
A downloaded model_final.pt file or its containing directory can be passed
to from_pretrained instead of the Hub repository ID. The checkpoint uses
PyTorch serialization and is loaded with weights_only=True.
License and attribution
Original DispViT contributions and project-created checkpoint weights are licensed under Apache 2.0. Included third-party components retain their upstream copyright notices and license terms; see NOTICE. Redistribute the checkpoint with the applicable license and attribution files.
Citation
If you find DispViT useful in your research, please cite our ICLR 2026 paper:
@inproceedings{guan2026dispvit,
title = {{DispViT}: Direct Stereo Disparity Regression with a Single-Stream Vision Transformer},
author = {Guan, Tongfan and Guo, Jiaxin and Huang, Tianyu and Dong, Jinhu and Wang, Chen and Liu, Yun-Hui},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2026},
url = {https://openreview.net/forum?id=c21yqwf02V}
}