ComfyUI - Technical Masterclass in AI Workflows

Get started with the most essential downloads

This lecture walks through the essential support files and tools required to get ComfyUI running securely and efficiently.

How rapid changes in generative AI lead to deprecation of resources, broken links, and student issues accessing models/content.

The course discusses what to do when resources disappear, the importance of reporting missing items, and how students can contact instructors for resolution.

Download the main Stable Diffusion 1.5 model from recommended site: Tensor Art for a safe and diverse set of models.  Rename the model to Dreamshaper_8.safetensor after download completes.

Many options = better reliability. 

Download the main Stable Diffusion 1.5 model from recommended site: Civit AI for a safe and very large selection of models.

Many options = better reliability. 

Install ComfyUI and get started on the journey

Nine months after the course began, Comfy.org began major structural changes to the software that have necessitated some updates to the courses.  You can find these further along in the course and this lecture discusses the course philosophy to handling these major updates.  The main purpose is to try to maintain the course relevance for as long as possible until eventually a new course becomes inevitable.

Discover how GPU classes, VRAM size, and architecture generations impact ComfyUI performance.

This lecture explains why certain cards outperform others and how to choose hardware that fits your workflow.

This lecture breaks down the RTX 50 Series Blackwell lineup, explaining memory bandwidth gains, tensor core improvements, new precision formats, and AI TOPS performance. You'll learn how each GPU fits different AI workloads and budgets.

Discover how modern GPU features such as GDDR7, FP8 precision, AI TOPS, and L2 cache impact your AI workflows. We also cover driver stability, long-term support, and how to build a smart upgrade path.

Explore the strengths and limitations of the RTX 40 Series alongside earlier models. We cover VRAM trends, launch history, Tensor Core improvements, and the product availability cycles that affect real-world pricing.

The ComfyUI Manager is a uniquely useful custom node.

Why you *may* want to consider learning to use Windows Powershell.

Updating with Git Pull

An introduction to most powerful custom node in ComfyUI.

Workflows are central to everything in ComfyUI.

How to customize your workflows.

Enjoy this practical exercise with workflows.

Workflow exercise outcome.

Learn about all the possibilities associated with workflows and ComfyUI files.

This lecture shows how ComfyUI workflows can be saved, restored, and recovered using built-in tools that many users overlook.

In this lecture, you’ll learn how to:

• Store complete workflows inside images

• Restore workflows from preview images, even when no Save Image node was used

• Work with multiple tabbed workflows inside a single ComfyUI session

• Save, export, and reopen workflows, and understand where ComfyUI stores them on disk

• Browse and load workflow templates as starting points for new projects

• Trace images back to their source nodes using built-in navigation tools

Why this matters:
These techniques help prevent lost work, speed up experimentation, and make workflows easy to reuse and share.

Leave with a clear “map” of the node chain you will master in the next lectures

1. Understand what “loading a model” means in ComfyUI: file selection → model object creation → usable outputs

2. Learn why SD 1.5 workflows depend on the three components: denoiser (MODEL), prompt encoder (CLIP), and decoder (VAE)

How prompts are converted into conditioning using CLIP Text Encode node.

Understand the role of tokens and embeddings in simple, practical terms that match what you see in the graph.

Develop a basic understanding of How CLIP Interprets Language: tokens, ambiguity, and prompt failure modes

You will develop insight into how CLIP tokenizations influence image generation in Stable Diffusion 1.5. The lecture highlights common prompt failure modes, including typos and semantic ambiguity, and explains why CLIP may prioritize unintended meanings (based on its training).

Using real examples, you explore how spelling errors, ambiguous phrases, and shared word meanings alter conditioning, and how careful wording, context, and negative prompts can reduce misinterpretation.

• Understand what a latent is, why Stable Diffusion operates in latent space, and how it is initialized

• Learn the conceptual role of seed, sampler, scheduler, and denoise in the sampling process

• Learn how ComfyUI separates latent creation from latent transformation

Rather than focusing on formulas, this lecture encourages you to feel what denoise does by running a simple experiment once and observing the results. You learn how denoise values shape image evolution.

Through a systematic hands-on test, you observe how lowering denoise changes image stability and generation time in a workflow of your own choice.  We build on this simple experiment in later lessons.

Explain what the VAE Decode node does and why it is required after sampling

Understand how latent samples from the KSampler are converted into a visible image

Clarify the difference between preview images and saved images in ComfyUI

Getting started with image to image workflows

Understanding more about image to image workflows.
Lecture Description: Image Resizing and Controlled Variation in ComfyUI

In this lecture, we explore practical image-to-image workflows in ComfyUI, focusing on resizing large images, controlling denoise, and understanding how prompts, steps, and models influence results.

Key topics covered:

• Reducing large images to manageable sizes using core ComfyUI scaling nodes

• Preserving structure while changing content

• Sampler and step experimentation

• Prompt structure and negative prompting

• Model limitations and model choice

  • Identifying when a model struggles due to limited training data

  • Why larger or newer models may perform better for specific domains, like food imagery

• Native vs custom-node workflows

Learn how to manage multiple installations of ComfyUI and how to point ComfyUI to key local installations including Automatic 1111 installations to access models, using the Extra Paths to Models Yaml file. 

Discover vital software to assist in managing all your files. 

If the extra paths to models yaml file included here doesn't work, it means that you need to edit a brand new version of the file that comes from your specific installation of ComfyUI.  This provided file has been tested with ComfyUI installations to upto December 2025

The ComfyUI Desktop Edition has a similar feature-set to the standalone version but installing it is a different experience.  Find out more about this version of ComfyUI, that can be installed independently.  Update September 2025

A quick tour of what you may find after installing the The ComfyUI Desktop Edition - if all went well. 

This module guides students through the latest improvements and features in ComfyUI, focusing on how these upgrades benefit practical workflows in Gen AI projects.

In this technical deep dive, we decode the specific batch file that acts as a "turbo button" for your ComfyUI workflows. We move beyond the basics to explain exactly how the run_nvidia_gpu_fast_fp16_accumulation.bat script modifies PyTorch's behavior, forcing it to abandon safety buffers in favor of raw speed. You will learn why this simple file swap can result in 30-40% faster generation times on NVIDIA RTX 40-series and 50-series cards by leveraging the full throughput of modern Tensor Cores.

ComfyUI's Nodes 2.0 beta represents a fundamental architectural shift that promises faster iterations and unlocked features previously impossible in the legacy interface. In this hands-on lecture, we activate the beta, compare visual changes, and systematically test critical workflows to identify what works—and what breaks.

You will learn how to run parallel installations using extra_model_paths.yaml for safe testing, explore the redesigned node layouts and theme options, and discover current limitations like missing seed controls and text scaling issues.

While not yet "prime time ready," understanding Nodes 2.0 now prepares you for the major UI evolution rolling out throughout 2026. This lecture serves as your technical roadmap to navigating the transition, complete with practical workarounds and beta testing strategies.

Master the process of installing, updating, and managing custom nodes using the new ComfyUI Manager interface. This lecture demonstrates the installation workflow.

Learn how to configure and switch between the legacy ComfyUI Manager and the new built-in manager during the transition period.  Back up your original batch files if you need to decide to use the edited ones below.

How to switch between old and new Manager in ComfyUI Desktop edition since January 2026 changes

This lecture explains how Dynamic VRAM in ComfyUI manages tensors and model weights to optimize memory usage. Learn how it reduces out-of-memory errors, improves loading times, and enables larger model execution. Ideal for building stable, high-performance AI workflows.

Purpose of lecture

To explain the internal mechanics and benefits of Dynamic VRAM in ComfyUI.

Why it matters technically

Dynamic VRAM minimizes inefficient memory swapping, reduces reliance on page files, and enables execution beyond traditional RAM limits.