MEt3R: Measuring Multi-View Consistency in Generated Images [CVPR 2025]

Mohammad Asim¹, Christopher Wewer¹, Thomas Wimmer^{1, 2}, Bernt Schiele¹, Jan Eric Lenssen¹

¹Max Planck Institute for Informatics, Saarland Informatics Campus, ²ETH Zurich

Project Page

TL;DR: A differentiable metric to measure multi-view consistency between an image pair.

📣 News

• 15.04.2025 - Updates:
  • Added optical flow-based warping backbone using RAFT.
  • Added psnr, ssim, lpips, rmse, and mse metrics on warped RGB images instead of feature maps.
  • Added nearest, bilinear and bicubic upsampling methods.
  • Refactored codebase structure.
• 26.02.2025 - Accepted to CVPR 2025 🎉!
• 10.01.2025 - Initial code releases.

🔍 Method Overview

MEt3R evaluates the consistency between images $\mathbf{I}_1$ and $\mathbf{I}_2$. Given such a pair, we apply DUSt3R to obtain dense 3D point maps $\mathbf{X}_1$ and $\mathbf{X}_2$. These point maps are used to project upscaled DINO features $\mathbf{F}_1$, $\mathbf{F}_2$ into the coordinate frame of $\mathbf{I}_1$, via unprojecting and rendering. We compare the resulting feature maps $\hat{\mathbf{F}}_1$ and $\hat{\mathbf{F}}_2$ in pixel space to obtain similarity $S(\mathbf{I}_1,\mathbf{I}_2)$.

📋 Contents

• 📓 Abstract
• 📌 Dependencies
• 🛠️ Quick Setup
• 📣 Example Usage
• 👷 Manual Install
• 📘 Citation

📓 Abstract

We introduce MEt3R a metric for multi-view consistency in generated images. Large-scale generative models for multi-view image generation are rapidly advancing the field of 3D inference from sparse observations. However, due to the nature of generative modeling, traditional reconstruction metrics are not suitable to measure the quality of generated outputs and metrics that are independent of the sampling procedure are desperately needed. In this work, we specifically address the aspect of consistency between generated multi-view images, which can be evaluated independently of the specific scene. Our approach uses DUSt3R to obtain dense 3D reconstructions from image pairs in a feed-forward manner, which are used to warp image contents from one view into the other. Then, feature maps of these images are compared to obtain a similarity score that is invariant to view-dependent effects. Using MEt3R, we evaluate the consistency of a large set of previous methods for novel view and video generation, including our open, multi-view latent diffusion model.

📌 Dependencies

- Python >= 3.6
- PyTorch >= 2.1.0
- CUDA >= 11.3
- PyTorch3D >= 0.7.5
- FeatUp >= 0.1.1

NOTE: Pytorch3D and FeatUp are automatically installed alongside MEt3R.

Tested with CUDA 11.8, PyTorch 2.4.1, Python 3.10

🛠️ Quick Setup

Simply install MEt3R using the following command inside a bash terminal assuming prequisites are aleady installed and working.

pip install git+https://github.com/mohammadasim98/met3r

💡 Example Usage

Simply import and use MEt3R in your codebase as follows.

import torch
from met3r import MEt3R

IMG_SIZE = 256

# Initialize MEt3R
metric = MEt3R(
    img_size=IMG_SIZE, # Default to 256, set to `None` to use the input resolution on the fly!
    use_norm=True, # Default to True 
    backbone="mast3r", # Default to MASt3R, select from ["mast3r", "dust3r", "raft"]
    feature_backbone="dino16", # Default to DINO, select from ["dino16", "dinov2", "maskclip", "vit", "clip", "resnet50"]
    feature_backbone_weights="mhamilton723/FeatUp", # Default
    upsampler="featup", # Default to FeatUP upsampling, select from ["featup", "nearest", "bilinear", "bicubic"]
    distance="cosine", # Default to feature similarity, select from ["cosine", "lpips", "rmse", "psnr", "mse", "ssim"]
    freeze=True, # Default to True
).cuda()

# Prepare inputs of shape (batch, views, channels, height, width): views must be 2
# RGB range must be in [-1, 1]
# Reduce the batch size in case of CUDA OOM
inputs = torch.randn((10, 2, 3, IMG_SIZE, IMG_SIZE)).cuda()
inputs = inputs.clip(-1, 1)

# Evaluate MEt3R
score, *_ = metric(
    images=inputs, 
    return_overlap_mask=False, # Default 
    return_score_map=False, # Default 
    return_projections=False # Default 
)

# Should be between 0.25 - 0.35
print(score.mean().item())

# Clear up GPU memory
torch.cuda.empty_cache()

Checkout example.ipynb for more demo examples!

👷 Manual Install

Additionally MEt3R can also be installed manually in a local development environment.

Install Prerequisites

pip install -r requirements.txt

Installing FeatUp

MEt3R relies on FeatUp to generate high resolution feature maps for the input images. Install FeatUp using the following command.

pip install git+https://github.com/mhamilton723/FeatUp

Refer to FeatUp for more details.

Installing Pytorch3D

MEt3R requires Pytorch3D to perform point projection and rasterization. Install it via the following command.

pip install git+https://github.com/facebookresearch/pytorch3d.git

In case of issues related to installing and building Pytorch3D, refer to Pytorch3d for more details.

Installing DUSt3R

At the core of MEt3R lies DUSt3R which is used to generate the 3D point maps for feature unprojection and rasterization. We adopt DUSt3R as a submodule which can be downloaded as follows:

git submodule update --init --recursive

📘 Citation

When using MEt3R in your project, consider citing our work as follows.

@inproceedings{asim24met3r,
    title = {MEt3R: Measuring Multi-View Consistency in Generated Images},
    author = {Asim, Mohammad and Wewer, Christopher and Wimmer, Thomas and Schiele, Bernt and Lenssen, Jan Eric},
    booktitle = {Computer Vision and Pattern Recognition ({CVPR})},
    year = {2024},
}