Skip to main content

Diffusion Models




What is the Diffusion model?

Diffusion models lie in the domain of Computer Vision. Diffusion is the model of deep learning that deals with latent or hidden variables in an image by adding or removing noise.

The diffusion model is a computational framework used in psychology and neuroscience to describe decision-making processes and response times in tasks involving perceptual discrimination or categorization. It proposes that decision-making involves accumulating evidence over time from noisy sensory inputs until a threshold is reached, at which point a decision is made.

What is noise?

Unwanted information that disrupts the clarity and accuracy of the picture.

the context of the diffusion model, "noise" refers to random variability or fluctuations in the accumulation of evidence over time. These fluctuations can arise from various sources, including variability in sensory inputs, neural noise, and random fluctuations in cognitive processing.





Stable Diffusion:

Stable diffusion is a type of Diffusion model used to generate text to image.

How Diffusion Models are made?

Diffusion models are made to make the image noisy, then, by removing that noise and getting back the original image.

In simpler terms, stable diffusion implies that the decision-making process has settled into a consistent pattern, and further accumulation of evidence does not significantly alter the outcome.

Why are Latent Diffusion models made?

In the beginning, Stable diffusion was used but, DM is very expensive and more computational. Therefore, Latent Diffusion Models are made. LDM converts the picture into latent(hidden)  representation (vector representation) due to which its computations decrease, and then apply diffusion model concept on it. LDM is 48 times faster than Diffusion( Pixel by Pixel).

How many types of Diffusion process?

Two types of Diffusion process are there;

  1. Forward Diffusion Process: In this process, noise is added into the image continuously.

  2. Reverse/ Backward Diffusion Process: In this process, noise is removed from the image until it gets the original image.  

How to add a noise in an image

There are two ways of adding noise into the picture.

  • By unit time:

When you gradually add noise to an image over time, it means you're slowly making the image more distorted. Instead of adding all the noise at once, you do it little by little. This helps you control how much the image changes. By adjusting how quickly you add the noise and how much you add each time, you can decide how blurry or messy you want the image to look. This method is useful when you want to carefully change an image without making it too messy."
  • By CoSin:

This method uses a special math trick, often using something called cosine transformations, to add fuzziness to a picture. By using this math trick, we can add fuzziness in a way that we can control, making the picture look different in the way we want it to.

Some popular AI Image Generator Tools:

Here are some popular AI image generator tools:

  1. DeepArt.io:

  2. DeepArt uses neural networks to apply artistic styles to images, allowing users to create unique artworks by combining different styles with their own photos.


  4. RunwayML:

  5. RunwayML offers a range of AI-powered tools for creatives, including image generation, style transfer, and image manipulation. It provides an intuitive interface for experimenting with various AI models.


  7. DeepDream Generator:

  8. DeepDream Generator uses Google's DeepDream algorithm to generate psychedelic and surreal images. Users can upload their own images and customize parameters to create unique visual effects.

    1. DALL-E:

    2. Developed by OpenAI, DALL-E is an AI model capable of generating images from textual descriptions. It can create novel and realistic images based on specific prompts provided by users.


    4. Artbreeder:

    5. Artbreeder allows users to blend and morph images using AI algorithms, creating new and unique artworks. It offers a range of features for exploring and manipulating visual content.


    7. GANPaint Studio:

    8. GANPaint Studio uses generative adversarial networks (GANs) to edit and manipulate images in various ways. Users can add or remove objects, change colors, and perform other modifications using AI-powered tools.


    10. DeepFaceLab:

    11. DeepFaceLab is a powerful tool for face swapping and facial reenactment using deep learning techniques. It enables users to create realistic face swaps in videos and images with advanced customization options.

  10. Here are some generic prompt with generated AI images examples. Use these prompts and change words or phrases to your desired output from AI tools. 
Prompt:
Image of 3D cartoonish backside of a beautiful girl with long and brown hair in pink shiny gown under the tree alone;
Negative prompt:
Blurriness and fade colors:





Prompt:
Draw an image of black and white scenery with waterfall.
Negative Prompt:
Nothing










Comments

Post a Comment

Popular posts from this blog

Machine Learning Vs Deep Learning

Basics of Machine Learning, Deep learning or Neural Networks:   Machine learning: Machine Learning is a great part of the Artificial intelligence field and is a popular way to execute AI projects through different methods. Normally in ML, there is no need to instruct computers, data is given to the system and according to prior experience and data, rules are made to apply further.  In the Machine Learning domain, models are trained on learning how to create rules for themselves to perform lookalike tasks by taking specific inputs by their desired possible outputs. ML typically requires processed or organized data with the involvement of humans. What is Deep Learning Study of neural networks is called deep learning.  DL is the study of doing ML with neural networks. In 2012, Neural Networks made up of basic and fundamental mathematical functions like addition, subtraction, multiplication and division etc. these networks  are like neurons in the human brain to make ...
1 comment
Read more

Prompt Engineering

  What is Prompt Engineering? Prompt Engineering: Prompt engineering is a way of asking a big AI system like GPT-4 to give  the answer we want. We have to be careful and clever about how to ask the question, because the AI system might not understand us or give us wrong or useless information. We can use different tricks to make our question better, like giving examples, rules, texts, or steps. Prompt engineering can help us talk to the AI system using normal language, without needing to know a lot of technical stuff. Prompt Engineering deals with the challenge of getting best possible outcomes utilizing token limits. What is the Anatomy of Prompt ? Anatomy of Prompt: In the context of language models, a prompt refers to the input or instructions provided to the model to generate the specific response or output. Characteristics of prompts are also called anatomy of prompts. Following step are needed to be taken while making a prompt; ☆Simulate persona: In this step ,first of ...
Post a Comment
Read more

Discriminative Artificial Intelligence Vs Generative Artificial Intelligence

  Discriminative Artificial Intelligence(AI): Discriminative Artificial Intelligence(AI) models are taught to distinguish between different data classes, patterns, images etc. In simple words it is a type of artificial intelligence that helps us tell different things apart. Discriminative AI focuses on learning the lines that separate different groups in our data. Instead of making new data, it looks at data we already have to figure out what it is.   Application Areas: Image Recognition: Discriminative AI helps identify what's in pictures, like telling if it's a cat or a dog. Speech Recognition: It figures out different words or phrases when people talk.  Natural Language Processing (NLP): It sorts text into different groups, like figuring out if a review is positive or negative. How Discriminative AI Works: Training Data: Discriminative AI learns from examples, like pictures of cats and dogs, to get better at telling them apart. Learning Boundaries: It figures ou...
1 comment
Read more