PixNerd Diffusers

Diffusers-style implementation of PixNerd (Pixel Neural Field Diffusion), following the layout of NiT-diffusers.

Legacy training stacks, spec-based model loaders, and bundled checkpoint pipelines have been removed. Inference uses native Diffusers components under src/diffusers.

Package layout

• src/diffusers/models/transformers/transformer_pixnerd.py — PixNerdTransformer2DModel and PixNerDiT
• src/diffusers/models/autoencoders/autoencoder_pixel.py — PixNerdPixelVAE
• src/diffusers/models/conditioners/conditioner_pixnerd.py — PixNerdLabelConditioner
• src/diffusers/schedulers/scheduling_flow_match_pixnerd.py — PixNerdFlowMatchScheduler
• src/diffusers/pipelines/pixnerd/pipeline_pixnerd.py — PixNerdPipeline
• scripts/convert_pixnerd_ckpt_to_diffusers.py — convert raw .ckpt files to a Diffusers pipeline directory
• scripts/sample_pixnerd.py — sample from a converted pipeline

Install

pip install -e .

Or install dependencies only:

pip install -r requirements.txt

Sample from CLI

python main.py sample \
  --pretrained_model_name_or_path path/to/converted-checkpoint \
  --class_label 207 \
  --num_images_per_prompt 4 \
  --output_dir samples

Sample script

python scripts/sample_pixnerd.py \
  --model path/to/converted-checkpoint \
  --class-label 207 \
  --height 512 \
  --width 512 \
  --num-inference-steps 25 \
  --guidance-scale 4.0 \
  --timeshift 3.0 \
  --order 2

Convert raw ImageNet checkpoints

python scripts/convert_pixnerd_ckpt_to_diffusers.py \
  --checkpoint raw/imagenet256/epoch%3D319-step%3D1600000_emainit.ckpt \
  --output models/BiliSakura/PixNerd-diffusers/PixNerd-XL-16-256

Batch conversion:

python scripts/convert_raw_imagenet_ckpts.py

Converted directories contain model_index.json, separate transformer/, scheduler/, vae/, and conditioner/ subfolders compatible with PixNerdPipeline.from_pretrained.

Python API

import sys
from pathlib import Path

sys.path.insert(0, str(Path("src").resolve()))
from diffusers import PixNerdPipeline

pipe = PixNerdPipeline.from_pretrained(
    "models/BiliSakura/PixNerd-diffusers/PixNerd-XL-16-512",
    torch_dtype=torch.bfloat16,
)
images = pipe(class_labels="golden retriever", num_inference_steps=25, guidance_scale=4.0).images

Gradio demo

python app.py --pretrained_model_name_or_path path/to/converted-checkpoint

Upstreaming to Diffusers

Copy the files under src/diffusers into the matching locations in the huggingface/diffusers repository and register the classes in Diffusers' lazy import tables.

Reference

@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},
}