MDPs and POMDPs in Julia - An interface for defining, solving, and simulating fully and partially observable Markov decision processes on discrete and continuous spaces.
Set of robotic environments based on PyBullet physics engine and gymnasium.
Key parameters and specifications
Key parameters and specifications
Parameter
Value
Language
Python
Category
Model Development
Subcategory
Domain-specific training
Keywords
Computer vision, Datasets, Image production
Stars
769
Forks
31
Contributors
1
Created date
2022-10-27
Updated date
2026-10-13
Stars in 1 day
0
Stars in 7 days
-2
Rank
14,012
Previous rank
13,888
Rank change
-124
7-day stars rank
18,431
Previous 7-day stars rank
18,365
7-day stars rank change
-66
Total stars rank
13,783
Previous total stars rank
13,761
Total stars rank change
-22
README summary
Color Diffusion Using diffusion models to colorize black and white images. Overview This project is a simple example of how we can use diffusion models to colorize black and white images. This implementation uses the LAB color space, a 3 channel alternative to the RGB color space. The "L" (Lightness) channel in this space is equivalent to a greyscale image: it represents the luminous intensity of each pixel. The two other channels are used to represent the color of each
pixel. To train the model, we first load color images and convert them to LAB. Then, we add noise only to the color channels, keeping the L channel constant. The model gets this channel "for free" because it doesn't need to learn how to predict the greyscale image: it is always known at train and test time. Forward Diffusion Process *Note that we actually don't need to go through all of the steps of the diffusion process to get to
timestep t. Our forward diffusion process is non-Markovian, but the entire diffusion process is shown for illustrative purposes* The model is a UNet that takes a 3 channel LAB input (the ground-truth greyscale channel concatenated with noised AB channels) and outputs a 2 channel prediction of the color noise. Forward diffusion and denoising at train time In addition to receiving the greyscale channel as input, the UNet is also conditioned on features extracted from the greyscale channel. Intermediate
feature maps from an encoder (implemented as the first half of a UNet) are concatenated with the features of the main denoising UNet throughout the downsampling stage of the forward pass. Future Work / Ideas This was just a quick proof of concept to satisfy my curiosity and get a feel for training diffusion models from scratch, so the results are very basic. There are many ways this project could be improved, such as: Using pretrained face recognition networks like ArcFace
or FaceNet as feature extractors to get the conditioning features Implementing cross attention on the embeddings Pretraining the greyscale feature extractor as the encoder stage of a greyscale autoencoder References A lot of code for the dataset and LAB color operations was adapted from moein-sharitania's colorization project, which used Conditional GANs I implemented optional dynamic thresholding as in Assembly AI's Minimagen project (the Assembly AI blog posts are excellent for getting a deep understanding of the maths and concepts
behind diffusion models) The UNet architecture was adapted from denoising-diffusion-pytorch Usage Run to download the CelebA dataset and extract it Use inference.
Frequently asked questions
How many stars and forks are recorded?
ErwannMillon/Color-diffusion has 769 stars and 31 forks.
What is the repository rank?
ErwannMillon/Color-diffusion is ranked 14,012.
What language and category are recorded?
ErwannMillon/Color-diffusion has 769 stars, uses Python and belongs to Model Development.
Which repository is closest in rank?
ErwannMillon/Color-diffusion is compared with UM-ARM-Lab/pytorch_mppi, ranked 14,014 with 769 stars.
README
Color Diffusion
Using diffusion models to colorize black and white images.
Overview
This project is a simple example of how we can use diffusion models to colorize black and white images.
This implementation uses the LAB color space, a 3 channel alternative to the RGB color space.
The "L" (Lightness) channel in this space is equivalent to a greyscale image: it represents the luminous intensity of each pixel. The two other channels are used to represent the color of each pixel.
To train the model, we first load color images and convert them to LAB.
Then, we add noise only to the color channels, keeping the L channel constant. The model gets this channel "for free" because it doesn't need to learn how to predict the greyscale image: it is always known at train and test time.
Forward Diffusion Process
Note that we actually don't need to go through all of the steps of the diffusion process to get to timestep t. Our forward diffusion process is non-Markovian, but the entire diffusion process is shown for illustrative purposes
The model is a UNet that takes a 3 channel LAB input (the ground-truth greyscale channel concatenated with noised AB channels) and outputs a 2 channel prediction of the color noise.
Forward diffusion and denoising at train time
In addition to receiving the greyscale channel as input, the UNet is also conditioned on features extracted from the greyscale channel. Intermediate feature maps from an encoder (implemented as the first half of a UNet) are concatenated with the features of the main denoising UNet throughout the downsampling stage of the forward pass.
Future Work / Ideas
This was just a quick proof of concept to satisfy my curiosity and get a feel for training diffusion models from scratch, so the results are very basic. There are many ways this project could be improved, such as:
Using pretrained face recognition networks like ArcFace or FaceNet as feature extractors to get the conditioning features
Implementing cross attention on the embeddings
Pretraining the greyscale feature extractor as the encoder stage of a greyscale autoencoder
