Denoising
Denoising is the process of removing unwanted noise or corruption from data while preserving meaningful structure.
Build a Text-to-Image Generator (from Scratch)
A denoising process is a procedure that progressively removes noise from a data sample to produce a clean result.
- How it is used
- Turns random noise into generated images, first unconditionally and then under text-prompt guidance.
- How deep it goes
- Works through forward and reverse time steps, noise prediction, and training; uses a library U-Net rather than implementing its architecture here.
- Examples
- Uses Fashion MNIST clothing images for generation and a motorcycle-riding banana prompt for latent text-to-image generation.
- What you can do after
- Apply it via diffusers to train a denoising U-Net and generate clothing-item images from random noise.
Read how Build a Text-to-Image Generator (from Scratch) explains denoising →
Grokking AI Algorithms, Second Edition
A denoising process is an iterative procedure that removes predicted noise from a corrupted representation to recover structured data.
- How it is used
- Turns random noise into a prompt-guided image and selectively fills masked regions when editing existing images.
- How deep it goes
- Traces training examples and the repeated predict-and-subtract procedure, including timestep scaling and noise-schedule trade-offs.
- Examples
- Follows a tree emerging from random noise and shows image progressions across denoising timesteps.
- What you can do after
- Evaluate/understand how a trained diffusion model generates or edits an image through repeated noise prediction and removal.
Read how Grokking AI Algorithms, Second Edition explains denoising →
Learn Generative AI with PyTorch
The process of removing noise from an image estimate one step at a time during generation.
- How it is used
- Generates clean images by iteratively reversing a diffusion process from random noise.
- How deep it goes
- Explains iterative prediction, U-Net structure, scheduler steps, and visualization, but does not derive the diffusion equations.
- Examples
- Demonstrates 50-step flower generation, fixed seeds, intermediate image storage, and visual comparisons across time steps.
- What you can do after
- Builds a denoising generation loop with a trained U-Net and DDIMScheduler, then visualizes its intermediate outputs.
Read how Learn Generative AI with PyTorch explains denoising →
Deep Learning with PyTorch, Second Edition
The process of removing noise from a corrupted input, here used iteratively in a diffusion model to reconstruct an image.
- How it is used
- Uses iterative noise removal to synthesize images, perform image-to-image generation, and constrain local edits through inpainting masks.
- How deep it goes
- Explains forward and reverse processes, controlled schedules, sampling steps, and PyTorch implementation while deferring the mathematical derivation.
- Examples
- Demonstrates horse-to-zebra inpainting, masked image editing, and PyTorch-logo generation from random points.
- What you can do after
- Apply it via iterative sampling to reconstruct data, generate images, or edit masked regions while preserving unmasked content.
Read how Deep Learning with PyTorch, Second Edition explains denoising →
Deep Learning with Python, Third Edition
- How it is used
- Uses denoising for image cleanup, super-resolution, and generating images from pure noise or text prompts.
- How deep it goes
- Explains autoencoder cleanup, iterative reverse diffusion, and text conditioning, but stops short of training a text-to-image model from scratch.
- Examples
- Uses low-quality JPEGs, smartphone super-resolution, pure-noise image generation, and Stable Diffusion prompts as examples.
- What you can do after
- Apply it via KerasHub Stable Diffusion to generate prompt-conditioned images and vary the number of denoising steps.
Read how Deep Learning with Python, Third Edition explains denoising →
Deep Learning with R, Third Edition
Denoising is the removal of unwanted or corrupting variation from data to recover a cleaner version.
- How it is used
- Uses noisy images as inputs for reconstruction, super-resolution, and image generation through repeated reverse diffusion.
- How deep it goes
- Explains the intuition and training sequence, including normalization, schedules, noise injection, reconstruction, and mask loss.
- Examples
- Uses low-quality JPEGs, smartphone super-resolution, blurred moon images, and pure-noise image generation as examples.
- What you can do after
- Build it from scratch by implementing a denoising model, diffusion schedule, reconstruction method, and noise-mask loss.
Read how Deep Learning with R, Third Edition explains denoising →
Geometry for Programmers
Denoising is the removal or reduction of unwanted variation and artifacts in sampled image data while preserving meaningful structure.
- How it is used
- Removes scanning and computation artifacts from binary CT-derived image and voxel models.
- How deep it goes
- Explains morphology, distance-based dilation, inversion-based erosion, and a simplified implementation without diagonal elements.
- Examples
- Uses binary 2D and 3D voxel models containing isolated noise, thin features, and thick features.
- What you can do after
- Apply it via binary voxel operations to remove small artifacts and smooth model boundaries.
Related concepts
- diffusion modelcontrasts with · A diffusion model is a generative model that learns to produce data by reversing a process that progressively adds noise to data.
- noiseaddresses · Noise is variation, content, or data treated as unwanted, irrelevant, or distinct from the signal, pattern, or structure of interest.
- erosionuses · Erosion is a morphological operation that shrinks foreground regions or, equivalently, produces an inward offset according to a structuring element.
- dilationuses · Dilation is a morphological operation that expands foreground regions or, equivalently, produces an outward offset according to a structuring element.
- embeddingused by · An embedding is a numerical vector representation of an item that encodes features or relationships useful for computational tasks.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.