VistaDream

This is the official PyTorch implementation of the following publication:

VistaDream: Sampling multiview consistent images for single-view scene reconstruction Haiping Wang, Yuan Liu, Ziwei Liu, Wenping Wang, Zhen Dong, Bisheng Yang ICCV 2025 Paper | Project-page(with Interactive DEMOs)

🔭 Introduction

🆕 News

• 2026-10-18: The journal version of VistaDream has been accepted by IEEE TPAMI. See moge3-lbd branch.
• 2025-06-26: VistaDream is accepted by ICCV 2025!
• 2024-10-23: Code, [project page], and [arXiv paper] are aviliable.

💻 Recommended installation and depth-model upgrade !

• VistaDream depends on several GPU, native, and third-party packages. To avoid installation errors and version conflicts, we strongly recommend using Codex or another coding agent to create the environment step by step, verify each dependency, and record any platform-specific fixes.

• If you have already installed the original VistaDream environment, you can ask an agent to update the depth modules, for instance, using MoGe-3 for depth estimation and LingBot-Depth for depth completion. This upgrade provides a substantial improvement in reconstruction quality. The corresponding interactive interface is available in the moge3-lbd branch for reference.

💻 Requirements

The code has been tested on:

• Ubuntu 20.04
• 1 * GeForce RTX 4090.

🔧 Installation

For complete installation instructions, please see INSTALL.md.

🚅 Pretrained model

VistaDream is training-free but utilizes pretrained models of several existing projects. To download pretrained models for Fooocus, Depth-Pro, Prior-Depth-Anything, SD-LCM, run the following command:

bash download_weights.sh

The pretrained models of LLaVA, MaskFormer, and Stable Diffusion-1.5 will be automatically downloaded from hugging face on the first running.

🔦 Generate your own scene (Interface)

Use vdplus conda environment, check INSTALL.md. Vistadream support interactive scene generation show above, please refer to INTERFACE.md.

🔦 Demo (Single-View Generation)

Try VistaDream using the following commands:

python demo.py

Then, you should obtain:

• data/sd_readingroom/scene.pth: the generated gaussian field;
• data/sd_readingroom/video_rgb(dpt).mp4: the rgb(dpt) renderings from the scene.

🔦 Demo (Sparse-View Generation)

To use sparse views as input as demo_here, we need VGGT to first reconstruct the input images to 3D as the scaffold (no zoom-out needed).

First download checkpoints and place it at tools/VGGT/checkpoints by the following command:

wget -P tools/VGGT/checkpoints https://huggingface.co/facebook/VGGT-1B/resolve/main/model.pt?download=true

Then try VistaDream with sparse images as input using the following commands:

python demo_sparse.py

Then, you should obtain:

• data/bedroom/scene.pth: the generated gaussian field;
• data/bedroom/video_rgb(dpt).mp4: the rgb(dpt) renderings from the scene.

🔦 Generate your own scene (Single or Sparse views as input)

If you need to improve the reconstruction quality of your own images, please refer to INSTRUCT.md

To visualize the generated gaussian field, you can use the following script:

import torch
from ops.utils import save_ply
scene = torch.load(f'data/vistadream/piano/refine.scene.pth')
save_ply(scene,'gf.ply')

and feed the gf.ply to SuperSplat for visualization.

🔦 ToDo List

• Early check in generation.
• Support more types of camera trajectory. Please follow Here to define your trajectory. An example is given in this issue.
• Support sparse-view-input (and no pose needed). An example is given in this issue.
• Interactive Demo.

🔗 Related Projects

We sincerely thank the excellent open-source projects:

• Fooocus for the wonderful inpainting quality;
• LLaVA for the wonderful image analysis and QA ability;
• Depth-Pro for the wonderful monocular metric depth estimation accuracy;
• Prior-Depth-Anything for the wonderful monocular metric depth completion accuracy;
• StableDiffusion for the wonderful image generation/optimization ability.