Beginners Guide to ComfyUI - 2026 Refresh with Seedance 2.0

By the end of this lecture, you will be able to:

• Run a simple ComfyUI workflow

• Understand the Comfy Cloud setup process for Comfy Online and for Comfy Desktop

• Navigate the infinite canvas using mouse controls

Use the .JSON File in the resources to open up this workflow

Alternatively use the templates section on the left to select the Kling: Image to Video template

Outcome:
You can run, navigate, and save your first ComfyUI workflow with confidence.

We continue with the previous video theme and learn some essential tricks to benefit from ComfyUI's infinite canvas.

By the end of this lecture, you will be able to:

• Adjust aspect ratios (9:16 vertical video)

• Understand how outputs are saved locally

• Save and reload ComfyUI workflows (JSON)

The attached video might be able to load the workflow which created it, or just view it for an example of the output.

In this lecture, you will learn how to:

• Use templates to start new workflows in ComfyUI

• Reopen workflows from disk

• Recreate workflows by dragging outputs onto the canvas

• Load videos and images into existing workflows

• Recognize when ComfyUI creates a workflow vs only loads media

Outcome:
You understand the difference between importing workflows and importing media in ComfyUI.

ComfyUI is developing at freight express speed, but it is easy to summarize and work with the key changes.

Don't miss this one, backwards compatibility is important.  This lecture is important.

In this lecture, you will learn how to:

• Change ComfyUI UI modes using the menu settings

• Switch between the legacy UI and the current UI

• Access the Manager and Manage Extensions

• Enable the Legacy Manager UI when needed

• Preview the Nodes 2.0 future mode safely

Outcome:
You can adapt when ComfyUI updates change where tools are located.

Purpose: Provide a focused review of Seedance 2.0, ByteDance's update to Seedance 1.5 Pro

Core technical focus: Video-audio latent space generation with inputs like text, images, reference videos, and PowerPoint slides.

Key models or systems involved: Compares to Sora 2.0 (poor aesthetics), Veo 3.1 (Google), Kling variants; excels in 15-second consistency.

Practical outcomes: Skills to critique cherry-picked demos for social media shorts, product viz, and edits like mirror inpainting.

Strategic relevance: Demonstrates how proprietary data (TikTok) outpaces open rivals, guiding decisions in video diffusion workflows.

Purpose of lecture

To introduce Seedance 2.0 capabilities and workflows within ComfyUI.

Why it matters technically

Seedance 2.0 enables high-fidelity multimodal video generation with temporal control, making it a leading model for production workflows.

Purpose of lecture

Systematically evaluate the results of a render and use the results to determine adjustments to process and settingsd

This lecture attemps a lo-fi integration of the default ComfyUI template with one of the inputs previously used and a look at the superior Seedance 2.0 full model.

This explores the surprisingly capable Text to Video capabilities of Seedance 2 which plays the model in a category of its own.

Lip syncing is a largely unadvertised strength in Seedance 2.0, but it arrives with a surprisingly annoying limitation in the ComfyUI API - and one which, unless it is soon addressed, will produce many irritating moments for beginner's who start using Seedance 2.0 in ComfyUI

This lecture is a turning point where ComfyUI moves beyond simple generation and into real creative control.
You will use reference images to guide motion, identity, and visual consistency, and turn still images into convincing video using the Kling O1 image-to-video model. This is the kind of workflow that starts to resemble real production, not just experimentation.

By the end of this lecture, you will be able to:

• Use ComfyUI templates to run Kling O1 image-to-video workflows

• Create video from reference images instead of text alone

• Combine prompts with visual references for better control

• Control aspect ratio, duration, and output format

• Improve facial consistency using multiple reference images

• Understand how cost and processing time change with settings

Outcome:
You can create controlled image-to-video results in ComfyUI using Kling O1, and understand how to guide motion and identity with references.

This lecture makes the creative process explicit by showing how prompt engineering is an iterative, hypothesis driven workflow inside ComfyUI, not a single shot instruction. You will see how models like Ideogram v3 exhibit selective attention to parts of a prompt, sometimes producing what can be described as catastrophic neglect of secondary concepts when multiple visual ideas compete for influence. Watching this unfold in real time helps you develop intuition for how generative systems prioritize signals and why creative control emerges from structured experimentation rather than from perfect prompts.

You will observe how changes to prompt order, seed control, and magic prompt affect the internal weighting of concepts, and how reference images used as creative anchors can shape ideation even when they are not directly fused into the output. This is the mindset shift from treating models as tools to treating them as stochastic collaborators with known failure modes.

By the end of this lecture, you will be able to:

• Run text-to-image workflows in ComfyUI using Ideogram v3

• Use seed locking and seed incrementing to study model behavior

• Diagnose prompt dominance and concept suppression

• Structure prompt order to influence composition

• Use magic prompt while understanding its transformation effects

• Iterate visually using inspiration images inside the canvas

Observed limitations in this method:

• Ideogram v3 can exhibit catastrophic neglect of secondary concepts

• Multi concept fusion is unreliable

• Magic prompt can amplify dominant motifs and suppress others

• Fixed seeds do not guarantee identical outputs across systems

Outcome:
You can design controlled text-to-image workflows in ComfyUI with Ideogram v3 and anticipate model failure modes during creative exploration.

This lecture drops you straight into large-scale model workflows. You will work with Qwen image diffusion models that are tens of gigabytes in size, far beyond what most local machines can comfortably run. Seeing these models inside Comfy Cloud makes the creative process tangible at a systems level. Creative control here is not about typing better prompts. It is about understanding how massive diffusion models, sampling pipelines, and workflow structure shape what is even possible to generate.

As you unpack subgraphs, compare full 50-step diffusion against LoRA accelerated pipelines, and run controlled experiments with fixed seeds, you will see non intuitive behavior. Smaller, faster adapters can sometimes outperform larger models in composition, while larger models may exhibit concept dominance and partial catastrophic neglect of secondary ideas. This lecture reframes creative work as applied model diagnostics.

By the end of this lecture, you will be able to:

• Run text-to-image workflows in Comfy Cloud using large Qwen image models

• Understand why model size and VRAM requirements change what you can do locally

• Unpack and edit complex workflows and subgraphs

• Compare LoRA accelerated pipelines with full diffusion pipelines

• Control seeds for fair model comparisons

• Interpret how K Sampler, CLIP text encoders, and latent sizes affect outputs

Limitations you will observe:

1. Faster LoRA pipelines trade detail for speed

2. Seed determinism varies across hardware

3. Cloud workflows introduce cost and latency constraints

This lecture helps you build confidence in the Comfy Cloud user interface by walking through how your assets, workflows, and results are actually organized over time. You will see how the cloud experience differs from local ComfyUI, why persistence changes how you experiment, and how reviewing past outputs helps you diagnose creative failures such as catastrophic neglect when models ignore part of a prompt.

You will also compare outputs from different pipelines, including 4 step LoRA workflows versus 50 step diffusion workflows, and learn how to read visual cues like contrast, softness, and texture to understand which pipeline produced which result. Finally, you will tour Graph Mode versus Simple Mode (beta) and learn how to export work so it can be reused locally.

By the end of this lecture, you will be able to:

• Navigate the Comfy Cloud UI and find generated assets

• Review past workflows and compare multiple outputs per run

• Distinguish 4 step LoRA results from 50 step diffusion results

• Switch between Graph Mode and Simple Mode (beta)

• Export workflows and download assets for local use

• Use past results to diagnose catastrophic neglect and prompt issues

This lecture shows the real-world difference between running workflows in Comfy Cloud versus running them locally in ComfyUI Desktop. You will learn how exported workflows move between environments, why you may see missing models locally, and how hardware requirements like VRAM and system memory determine whether a workflow will run at all.

By the end of this lecture, you will be able to:

• Export and move ComfyUI workflows between Comfy Cloud and desktop

• Identify and respond to missing models prompts and dependency dialogs

• Estimate whether a workflow is “local-friendly” (10–40GB) vs “cloud-scale” (50+ GB)

• Understand tradeoffs: download size, VRAM, compute time, and electricity cost

• Use bypass to isolate what is failing inside a complex workflow

Outcome:
You can choose the right environment (cloud vs local) and avoid workflows your hardware cannot realistically support.

If you share or download workflows, custom nodes are often the reason they break. This lecture shows how workflow portability really works in ComfyUI, how to identify missing components, and how the Manager can install the missing custom nodes so the workflow runs again.

By the end of this lecture, you will be able to:

• Diagnose why a transferred workflow fails on a new machine

• Use the Manager to locate and install missing custom nodes

• Understand the difference between ComfyUI core nodes and third-party nodes

• Restart ComfyUI correctly after installing nodes so changes apply

Limitations you will encounter:

• Some workflows depend on specific versions of nodes or models

• UI changes can alter where options appear in the Manager

Outcome:
You can move workflows between machines without losing functionality.

This lecture walks through the complete integration of Google Nano Banana 2, also known as Gemini 3.1 Flash, within ComfyUI Desktop.

You will learn how to:

• Update ComfyUI Desktop safely
• Enable and configure the partner node
• Resolve template conflicts and startup dependency issues
• Scale outputs from 1K to 4K resolution
• Reuse seeds to preserve composition
• Shift from painterly realism to photographic rendering
• Test lighting transitions while maintaining scene structure
• Evaluate prompt adherence in complex multi-condition prompts

By the end of this lecture, you will understand how Flash-tier models behave under high-resolution workflows and how to maximise stylistic control using seed stability.

By completing this lecture, you will be able to:

• Confidently integrate the exceptional quality of Recraft 4 Pro Model into a ComfyUI workflow
• Generate professional poster-style and editorial layouts with typography, and realistic photography
• Diagnose aspect ratio limitations and plan around them