Mastering Diffusion Models: A Guide to Image Generation with Stable Diffusion

Introduction

Diffusion models, rooted in probabilistic generative modeling, are powerful tools for data generation. Initially in machine learning research, their history dates back to the mid-2010s when Denoising Autoencoders were developed. Today, they have gained prominence for their ability to generate high-quality images from text by modeling the denoising process. Current usage is in image synthesis, text generation, anomaly detection, finding utility in art, natural language processing, and cybersecurity. The future scope of diffusion models holds the potential for revolutionizing content creation, improving language understanding, making them a pivotal part of AI technologies, and solving real-world challenges. In this article, we will understand the basics of the diffusion model. Our focus will be on latent diffusion models related to text-to-image generation. We will learn to use image generation with the diffusion model in Python the Stable Diffusion model by Dream Studio. So let’s get started!

Learning Objectives

In this article, we will learn about

• Get an understanding of Diffusion models and their basics
• We will know about the architecture of Diffusion Models
• Get to know about the open-source diffusion model Stable Diffusion.
• We will learn to use Stable Diffusion for image generation using text in Python

This article was published as a part of the Data Science Blogathon.

Overview of Diffusion Models

Diffusion models belong to the class of generative models, meaning they can generate data similar to the one on which they are trained. In essence, the diffusion models destroy training data by adding noise and then learning to recover the training data by removing the noise. In the process, it learns the parameters of the neural network. We can then use this trained model and generate new data similar to training data by randomly sampling noise through the learned denoising process.  This concept is similar to Variational Autoencoders (VAEs) in which we try to optimize a cost function by first projecting the data onto the latent space and then recovering it back to the starting state. In diffusion models, the system aims to model a series of noise distributions in a Markov Chain and “decodes” the data by undoing/denoising the data in a hierarchical fashion.

Do you know the Basics of Diffusion Models?

A diffusion denoising process modeling basically involves 2 major steps  – the forward diffusion process (adding noise) and reverse diffusion process (removing noise). Let us try to understand each step one by one.

Forward Diffusion

The below are the steps for forward diffusion:

• The image(x0) is slowly corrupted iteratively in a Markov chain manner by adding scaled Gaussian noise.
• This process is done for some T time steps where we get xT.
• No model is involved during this step
• After this stage of Forward diffusion we have an image xT which is have Gaussian distribution. We have converted the data distribution into standard normal distribution with uniform variance.

Backward/ Reverse Distribution

• In this process we undo the forward diffusion and our objective is to remove the noise iteratively using a neural network model.
• The model’s task is to predict the noise added in image xt in time step t to image xt-1 . The model thus, predicts the amount of noise added in each time step to each sequence of images.

What is Stable Diffusion Framework?

Many open-source contributors collaborated to create the Stable Diffusion model, which is one of the most popular and efficient diffusion models available. It runs seamlessly on limited compute resources. It’s architecture consists of 4  components :-

1. Variational Autoencoders (VAE): Utilise it to decode pictures and translate them from latent space into pixel space. The latent space is a condensed representation of a picture that highlights its key elements. Working with latent embeddings is computationally lot cheaper and compress the latent spaces (have significantly lower dimensionality).

2. Text encoder and Tokenizer: To encode the user specific text prompt which is to generate the image.

3.  The U-Net Model: Latent image representations are denoised using it. Like an autoencoder, a U-Net has a contracting path and an expanding path. A U-Net does, however, have skip connections. These aid in the information propagation from the prior layers, which helps to solve the issue of disappearing gradients. Additionally, since we ultimately lose information in the contractive path, it aids in maintaining the finer details.

How to Use Stable Diffusion in Python for Image Generation?

In the below python implementation we will use the stable diffusion model to generate images.

1. Installing Libraries

!pip install transformers diffusers accelerate
!pip install xformers

2. Importing Libraries

from diffusers import StableDiffusionPipeline
import torch

3. Loading Stable Diffusion Model

Here we load the specific stable diffusion model in model_id below which is on Hugging face library.

model_id = "dreamlike-art/dreamlike-photoreal-2.0"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

4. Generate Prompts for Image

Here we generate 3 prompts for images we create 2 images of Alice in Wonderland with different styles and a third image of chesire cat.

prompts = ["Alice in Wonderland, Ultra HD, realistic, futuristic, detailed, octane render, photoshopped, photorealistic, soft, pastel, Aesthetic, Magical background",
          "Anime style Alice in Wonderland, 90's vintage style, digital art, ultra HD, 8k, photoshopped, sharp focus, surrealism, akira style, detailed line art",
          "Beautiful, abstract art of Chesire cat of Alice in wonderland, 3D, highly detailed, 8K, aesthetic"]


images = []

5. Save Images in the folder

for i, prompt in enumerate(prompts):
    image = pipe(prompt).images[0]
    image.save(f'picture_{i}.jpg')
    images.append(image)

Output Generated Images

Conclusion

In the realm of AI, researchers are currently exploring the powerful potential of diffusion models for wider application across various domains. Product designers and illustrators are experimenting with these models to quickly generate innovative prototype designs. Furthermore, several other robust models exist for generating more detailed images and can find utility in various photography tasks. Experts believe that these models will have a pivotal role in generating video content for influencers in the future.

Key Takeaways

• We understood the basic concepts behind diffusion models and their working principle.
• Stable diffusion is an important open source model and we learnt about its internal architecture.
• We learned how to run a stable diffusion model in Python to generate images using it with prompts.

Frequently Asked Questions

The media shown in this article is not owned by Analytics Vidhya and is used at the Author’s discretion.