Diffusion Model Architecture Slide and Detail
|
Diffusion models are a type of generative model that are trained to gradually denoise a noisy input image. This is done by starting with a completely noisy image and then gradually adding detail to the image until it is a realistic image. Diffusion models are trained using a technique called contrastive learning. In contrastive learning, the model is trained to distinguish between real images and images that have been generated by the model. The model is rewarded for correctly identifying real images and penalized for incorrectly identifying generated images. As the model is trained, it learns to gradually add detail to the noisy image until it is a realistic image. This is because the model learns to identify the patterns that are present in real images and to add these patterns to the noisy image. Diffusion models have been shown to be effective at generating high-quality images. They are also relatively fast to train, which makes them a promising approach for generating images in real time. Here is a more detailed description of the diffusion modeling architecture: Noisy image: The input to the diffusion model is a noisy image. The noise can be added to the image in a variety of ways, such as by adding Gaussian noise or by adding salt-and-pepper noise. Encoder: The encoder is a neural network that takes the noisy image as input and outputs a latent representation of the image. The latent representation is a high-dimensional vector that contains information about the important features of the image. Decoder: The decoder is a neural network that takes the latent representation as input and outputs a denoised image. The decoder uses the information in the latent representation to add detail to the noisy image until it is a realistic image. Contrastive loss: The contrastive loss is a loss function that is used to train the diffusion model. The contrastive loss is calculated by comparing the denoised image to the real image. The model is rewarded for correctly identifying the real image and penalized for incorrectly identifying the denoised image. |