metadata
license: mit
tags:
- diffusers
- ddpm
- unconditional-image-generation
- landscape
library_name: diffusers
pipeline_tag: unconditional-image-generation
ddpm-landscape
A 256x256 unconditional DDPM that generates natural landscape images. Full fine-tune of google/ddpm-church-256 on the Landscapes HQ (LHQ) dataset.
Usage
# !pip install diffusers
from diffusers import DDPMPipeline, DDIMPipeline, PNDMPipeline
model_id = "crab27/ddpm-landscape"
# load model and scheduler
ddpm = DDPMPipeline.from_pretrained(model_id) # you can replace DDPMPipeline with DDIMPipeline or PNDMPipeline for faster inference
# run pipeline in inference (sample random noise and denoise)
image = ddpm().images[0]
# save image
image.save("ddpm_generated_image.png")
Base model
google/ddpm-church-256— original 256x256 DDPM by Ho et al. All credit for the base architecture and pretrained weights goes to the original authors.
Dataset
- Landscapes HQ (LHQ), 256x256 split, from ALIS — Aligning Latent and Image Spaces to Connect the Unconnectable (Skorokhodov et al., 2021).
- Project / data: https://github.com/universome/alis
@article{ALIS,
title = {Aligning Latent and Image Spaces to Connect the Unconnectable},
author = {Skorokhodov, Ivan and Sotnikov, Grigorii and Elhoseiny, Mohamed},
journal = {arXiv preprint arXiv:2104.06954},
year = {2021}
}
Fine-tuning
| Base model | google/ddpm-church-256 |
| Dataset | LHQ (256x256) |
| Epochs | 50 |
| Batch size | 32 |
| Optimizer | AdamW |
| Learning rate | 1e-5 (cosine schedule, 500 warmup steps) |
| Loss | MSE on predicted noise |
| Augmentation | Random horizontal flip |
Acknowledgements
- Base weights: Google / the original DDPM authors (Ho, Jain, Abbeel, 2020).
- Dataset: Skorokhodov et al., authors of ALIS / LHQ — https://github.com/universome/alis
- Built with
diffusers.