Downloads · 30 days
41
49% of all-time downloads
Hannah1102/GAD-PixArt-XL-2-512x512
GAD-PixArt-XL-2-512x512 is a text-to-image model from Hannah1102. Use it when you need an image from a text prompt. It is set up for diffusers. The card lists the license as openrail++.
Official distilled PixArt checkpoint for:
Downloads · 30 days
41
49% of all-time downloads
All-time downloads
83
Public
Parameters
611M
2.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors2.4 GB · 100%
From the Hugging Face model README
Official distilled PixArt checkpoint for:
Restoring Initial Noise Sensitivity in Text-to-Image Distillation via Geometric Alignment
ICML 2026
Paper: https://arxiv.org/abs/2606.01651
Code: https://github.com/Hannah1102/GAD
This checkpoint is obtained by applying Geometry-Aware Distillation (GAD) to PixArt-XL-2-512x512.
PixArt-alpha/PixArt-XL-2-512x512
This repository contains the GAD-distilled Transformer weights based on PixArt-XL-2-512x512.
The tokenizer, text encoder, VAE, and scheduler are loaded from the original PixArt-XL-2-512x512 model, while the Transformer is replaced with the GAD-distilled checkpoint.
import torch
from diffusers import PixArtAlphaPipeline, Transformer2DModel
device = "cuda"
dtype = torch.bfloat16
base_model = "PixArt-alpha/PixArt-XL-2-512x512"
gad_model = "Hannah1102/GAD-PixArt-XL-2-512x512"
# Load the GAD-distilled Transformer
transformer = Transformer2DModel.from_pretrained(
gad_model,
torch_dtype=dtype,
)
# Load the remaining components from the original PixArt model
pipe = PixArtAlphaPipeline.from_pretrained(
base_model,
transformer=transformer,
torch_dtype=dtype,
use_safetensors=True,
)
pipe = pipe.to(device)
prompt = "A cinematic photo of an astronaut riding a horse on Mars"
generator = torch.Generator(device=device).manual_seed(42)
image = pipe(
prompt=prompt,
num_inference_steps=4,
guidance_scale=0.0,
generator=generator,
).images[0]
image.save("gad_pixart.png")
@inproceedings{huang2026restoring,
title={Restoring Initial Noise Sensitivity in Text-to-Image Distillation via Geometric Alignment},
author={Huang, Huayang and Wang, Ruoyu and Zhao, Jinhui and Deng, Wei and Zhou, Daiguo and Luan, Jian and Wu, Yu and Zhu, Ye},
booktitle={Proceedings of the 43rd International Conference on Machine Learning},
year={2026},
organization={PMLR}
}