PixNerd-XL-P16-C2I

PixNerd: Pixel Neural Field Diffusion

<div style="text-align: center;"> <a href="https://huggingface.co/papers/2507.23268"><img src="https://img.shields.io/badge/Paper-2507.23268-b31b1b.svg" alt="Paper"></a> <a href="https://github.com/MCG-NJU/PixNerd"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="Code"></a> <a href="https://huggingface.co/spaces/MCG-NJU/PixNerd"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Online_Demo-green" alt="Demo"></a> </div>

PixNerd is a novel pixel-space diffusion transformer for image generation, introduced in the paper PixNerd: Pixel Neural Field Diffusion. Unlike conventional diffusion models that depend on a compressed latent space shaped by a pre-trained VAE, PixNerd proposes to model patch-wise decoding with a neural field. This results in a single-scale, single-stage, efficient, and end-to-end solution that directly operates in pixel space, avoiding accumulated errors and decoding artifacts.

<p align="center"> <img src="https://huggingface.co/MCG-NJU/PixNerd/resolve/main/figs/arch.png" alt="PixNerd Architecture Diagram" width="700"> </p>

✨ Key Highlights

• Efficient Pixel-Space Diffusion: Directly models image generation in pixel space, eliminating the need for VAEs and their associated complexities or artifacts.
• Neural Field Decoding: Employs neural fields for patch-wise decoding, improving the modeling of high-frequency details.
• Single-Stage & End-to-End: Offers a simplified, efficient training and inference paradigm without complex cascade pipelines.
• High Performance: Achieves competitive FID scores on ImageNet 256x256 (2.15 FID) and 512x512 (2.84 FID) for unconditional image generation.
• Text-to-Image Extension: The framework is extensible to text-to-image applications, achieving strong results on benchmarks like GenEval (0.73 overall score) and DPG (80.9 overall score).

Visualizations

Below are sample images generated by PixNerd, showcasing its capabilities:

<p align="center"> <img src="https://huggingface.co/MCG-NJU/PixNerd/resolve/main/figs/pixelnerd_teaser.png" alt="PixNerd Teaser" width="700"> <br/> <img src="https://huggingface.co/MCG-NJU/PixNerd/resolve/main/figs/pixnerd_multires.png" alt="PixNerd Multi-Resolution Examples" width="700"> </p>

Checkpoints

The following checkpoints are available:

Online Demos

You can try out the PixNerd-XXL/16 (text-to-image) model on our Hugging Face Space demo: https://huggingface.co/spaces/MCG-NJU/PixNerd.

To host a local Gradio demo for text-to-image applications, run the following command after setting up the environment:

python app.py --config configs_t2i/inference_heavydecoder.yaml  --ckpt_path=XXX.ckpt

Usage

For image generation (C2i for ImageNet), you can use the provided codebase. First, install the required dependencies:

# for installation
pip install -r requirements.txt

Then, run inference using the main.py script (replace XXX.ckpt with your checkpoint path):

# for inference
python main.py predict -c configs_c2i/pix256std1_repa_pixnerd_xl.yaml --ckpt_path=XXX.ckpt
# or specify the GPU(s) to use:
CUDA_VISIBLE_DEVICES=0,1, python main.py predict -c configs_c2i/pix256std1_repa_pixnerd_xl.yaml --ckpt_path=XXX.ckpt

For more details on training and evaluation for both C2i and T2i applications, please refer to the official GitHub repository.

Citation

If you find this work useful for your research, please cite our paper:

@article{2507.23268,
Author = {Shuai Wang and Ziteng Gao and Chenhui Zhu and Weilin Huang and Limin Wang},
Title = {PixNerd: Pixel Neural Field Diffusion},
Year = {2025},
Eprint = {arXiv:2507.23268},
}

Acknowledgement

The code is mainly built upon FlowDCN and DDT.