Practical overview
When it fits
Use it when you want to explore a code-based workflow that infers a 3D object from one image and can manage the project’s model and package dependencies.
Who it is for
- Researchers studying single-image 3D generation and 3D vision
- Developers prototyping image-to-3D creation workflows with research code
Documented uses
- Generate a 3D object from one reference image.
- Progress from frontal-view optimization to full 360-degree geometry.
- Refine a coarse model into textured point clouds.
Requirements & limits
- The README pins PyTorch 1.10.0, torchvision 0.11.1 and torchaudio 0.10.0 for CUDA 11.3.
- The training workflow depends on DPT depth estimation, SAM foreground masks, BLIP2 captioning and a Stable Diffusion 2.0 prior.
- A Hugging Face token or CLI login may be required to access the pretrained diffusion model.
- The documented process is a training pipeline rather than a described hosted or one-click service.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the documented training environment
Install the listed Python packages and prepare the model dependencies used by the training workflow.
Before you start
- Python environment capable of installing the listed packages
- PyTorch 1.10.0, torchvision 0.11.1 and torchaudio 0.10.0 CUDA 11.3 packages
- Stable Diffusion 2.0 access; a Hugging Face token or CLI login may be needed
- DPT, SAM, BLIP2 and the listed project dependencies
Steps
Install the pinned PyTorch packages for CUDA 11.3.
Install the listed tiny-cuda-nn, CLIP, Diffusers, Hugging Face Hub, PyTorch3D and contextual-loss packages.
Install the repository requirements and local raymarching package.
Set up DPT and place the downloaded dpt_hybrid model file in dpt_weights; prepare SAM, BLIP2 and Stable Diffusion access.
Expected outcome
The documented dependencies and model resources are prepared for the training stages.
Guide limitations
- The Stable Diffusion model may require Hugging Face access credentials.
Run the documented two-stage image-to-3D training
Use a reference image for the coarse 360-degree geometry stage, then run the documented refinement stage.
Before you start
- Prepared dependencies and model resources
- Reference image path and workspace name
Steps
Run the coarse stage with the reference image and frontal camera range: python main.py --workspace NAME --ref_path IMGPATH --phi_range 135 225 --iters 2000.
Expand camera sampling to 360 degrees with: python main.py --workspace NAME --ref_path IMGPATH --phi_range 0 360 --albedo_iters 3500 --iters 5000 --final.
Run refinement on the coarse result with: python main.py --workspace NAME --ref_path IMGPATH --phi_range 135 225 --refine.
Expected outcome
The documented workflow progresses from a coarse full-view 3D result to a refined result.
Guide limitations
- Replace NAME and IMGPATH with the intended workspace name and reference-image path.
- The README notes that a back-view prompt can be requested with --need_back.
- The documented commands describe training; they do not specify a separate export workflow.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- junshutang/Make-It-3D · Official homepage title · retained in its source language