Practical overview
When it fits
Use NiT when you need a research codebase for multi-resolution, multi-aspect-ratio diffusion image synthesis and are prepared to work with local Python dependencies, checkpoints and datasets.
Who it is for
- Researchers studying diffusion models and image synthesis.
- Engineers seeking a codebase for experimenting with generation across image sizes and aspect ratios.
- Teams reproducing or evaluating class-guided ImageNet generation workflows.
Documented uses
- Generate class-guided images using the provided NiT-XL sampling scripts.
- Evaluate generated samples with image-generation metrics using the referenced ADM evaluation suite.
- Prepare ImageNet-1K data for training, including preprocessing and packing images with different resolutions.
Requirements & limits
The code is public on GitHub under the Apache-2.0 license. The repository provides links to pretrained models and processed ImageNet-1K data.
- The documented setup specifies Python 3.10, PyTorch 2.5.1, torchvision 0.20.1, CUDA 11.8 wheels and FlashAttention.
- Sampling scripts are listed for 256×256, 512×512 and 768×768; the README also reports benchmark results for other resolutions and aspect ratios.
- Training preparation includes downloading or locally preprocessing ImageNet-1K data and packing images with different resolutions.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the NiT Python environment
Install the repository and its documented Python dependencies as preparation for using its sampling and training workflows.
Before you start
- Git
- Conda
- Python 3.10
- PyTorch 2.5.1 and torchvision 0.20.1 using the CUDA 11.8 wheel index
- FlashAttention
Steps
Clone the NiT repository and change into its directory.
Create and activate a Conda environment named nit_env with Python 3.10.
Install the specified PyTorch and torchvision versions, FlashAttention, and the packages in requirements.txt.
Install the repository in editable mode with pip install -e .
Expected outcome
The repository is installed in the environment described by its setup instructions.
Guide limitations
- This setup recipe does not cover downloading model checkpoints, generating samples, or training.
- The README lists CUDA 11.8 wheel installation but does not specify hardware requirements.
Prepare NiT-XL checkpoints for sampling
Download the two NiT-XL checkpoint files listed in the repository model zoo.
Before you start
- The NiT repository cloned locally
- wget
- Network access to the listed Hugging Face model files
Steps
Create a checkpoints directory in the repository.
Use the README's wget commands to download model_1000K.safetensors and model_1500K.safetensors into that directory.
Expected outcome
The two checkpoint files are placed in the repository's checkpoints directory for the documented model workflows.
Guide limitations
- The listed sampling scripts cover 256×256, 512×512 and 768×768; this guide does not document sampling execution.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- WZDTHU/NiT · Official homepage title · retained in its source language