We propose PixNerd, a powerful and efficient pixel-space diffusion transformer for image generation (without VAE). Different from conventional pixel diffusion models, we employ the neural field to improve the high frequercy modeling .

• We achieve 1.93 FID on ImageNet256x256 Benchmark with PixNerd-XL/16 (1600k training steps).
• We achieve 2.84 FID on ImageNet512x512 Benchmark with PixNerd-XL/16.
• We achieve 0.73 overall score on GenEval Benchmark with PixNerd-XXL/16.
• We achieve 80.9 avergae score on DPG Benchmark with PixNerd-XXL/16.

We provide online demos for PixNerd-XXL/16(text-to-image) on HuggingFace Spaces.

强烈建议本地部署玩玩，线上的模型推理速度会慢一些。以及因为这个我把任意分辨率和动画都关了。

HF spaces: https://huggingface.co/spaces/MCG-NJU/PixNerd

To host the local gradio demo, run the following command:

# for text-to-image applications
python app.py --config configs_t2i/inference_heavydecoder.yaml  --ckpt_path=XXX.ckpt

For C2i(ImageNet), We use ADM evaluation suite to report FID.

# for installation
pip install -r requirements.txt

# 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 with as :
CUDA_VISIBLE_DEVICES=0,1, python main.py predict -c configs_c2i/pix256std1_repa_pixnerd_xl.yaml --ckpt_path=XXX.ckpt

# for training
# train
python main.py fit -c configs_c2i/pix256std1_repa_pixnerd_xl.yaml

For T2i, we use GenEval and DPG to collect metrics.

@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},
}

The code is mainly built upon FlowDCN and DDT.