ldm_cc3m_clean_condnoise_uniform Model Card

This repository contains diffusion models (DMs) developed for exploring the impact of slight corruption in pre-training data on generative performance. By introducing controlled condition corruption, we observed significant improvements in the quality, diversity, and fidelity of generated outputs across various DM architectures. This ldm_cc3m_clean_condnoise_uniform is trained on CC3M with CEP-U proposed in our paper.

The models and resources provided here aim to advance research in robust generative modeling and inspire new approaches to data-centric AI development. Explore the models at https://huggingface.co/DiffusionNoise.

Note: Since the models are trained on CC3M, which is a relatively small dataset, they might be incapable of follow complex prompts.

Usage

Currently, we only support a custon diffusers version. You need to install the diffusers from here: https://github.com/Hhhhhhao/diffusers.

Install

git clone https://github.com/Hhhhhhao/diffusers.git
cd diffusers
pip install -e ./

or 

pip install git+https://github.com/Hhhhhhao/diffusers.git

Use with (Custom) Diffusers

from diffusers.pipelines.latent_diffusion.pipeline_latent_diffusion import LDMTextToImagePipeline
from diffusers DPMSolverMultistepScheduler

# load model
pipeline = LDMTextToImagePipeline.from_pretrained('DiffusionNoise/ldm_cc3m_clean_condnoise_uniform')    
# use DPM scheduler for faster inference
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
# use float 16
pipeline = pipeline.to('cuda')
pipeline = pipeline.to(torch.float16)

# inference 
prompt = 'A photograph of a modern kitchen with appliances'
batch_size = 1
num_inference_steps=25
guidance_scale=2.5
images = pipeline([prompt] * batch_size, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale).images

images[0].save("img.png")

More comprehensive usage scripts can be found here: https://github.com/Hhhhhhao/DiffusionNoise

Direct Use

The model is intended for research purposes only. Possible research areas and tasks include:

• Safe deployment of models which have the potential to generate harmful content.
• Probing and understanding the limitations and biases of generative models.
• Generation of artworks and use in design and other artistic processes.
• Applications in educational or creative tools.
• Research on generative models.
• Excluded uses are described below.

Misuse, Malicious Use, and Out-of-Scope Use

The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.

The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.

Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:

• Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
• Intentionally promoting or propagating discriminatory content or harmful stereotypes.
• Impersonating individuals without their consent.
• Sexual content without consent of the people who might see it.
• Mis- and disinformation
• Representations of egregious violence and gore
• Sharing of copyrighted or licensed material in violation of its terms of use.
• Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.

Citation

@article{chen2024slight,
  title={Slight Corruption in Pre-training Data Makes Better Diffusion Models},
  author={Chen, Hao and Han, Yujin and Misra, Diganta and Li, Xiang and Hu, Kai and Zou, Difan and Sugiyama, Masashi and Wang, Jindong and Raj, Bhiksha},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2024}
}