ComfyUI concepts
An encyclopedia of ComfyUI for a newcomer, one idea per page. New here? Read Start here first: it orders these pages and pairs them with the lessons.
Basics
- Start here: The pages in the order a newcomer needs them.
- What ComfyUI is: The server, the editor, the queue, output nodes and where results appear.
- Nodes and links: How a graph is made of nodes joined by links.
- Data types: MODEL, CLIP, VAE, LATENT, IMAGE, CONDITIONING and others.
- Workflow formats: The workflow file, the API format and pictures with a workflow inside.
- Where ComfyUI files live: models/, input/, output/ and custom_nodes/.
Making a picture
- Text to image: The basic seven-node graph.
- Checkpoint: The main model file.
- Separate loaders: Diffusion model, text encoder and VAE as separate files.
- Text encoders: CLIP, T5 and why some models have two.
- Conditioning: The encoded prompt the sampler follows.
- The prompt: How models read text: tags or sentences, order, length.
- Prompt syntax and weights: What ComfyUI understands in a prompt, and what it does not.
- The negative prompt: Saying what you do not want.
- Latent: The compressed picture the model works on.
- VAE: Turning a latent into a picture and back.
- Sampler: The node that makes the picture step by step, and the sampler names.
- Scheduler: How the noise is removed over the steps.
- Steps: How many steps and what more of them changes.
- CFG: How strictly the model follows the prompt; Flux guidance.
- Seed: The number that makes a result repeatable.
- Denoise: How much of the starting picture is redrawn.
- Resolution: Picture sizes each family likes.
- Batch: Several pictures in one run.
Picture techniques
- Picture to picture: Starting from a picture instead of noise.
- Inpainting: Redrawing only part of a picture, and masks.
- Outpainting: Extending a picture beyond its edges.
- LoRA: A small add-on that changes style or subject.
- ControlNet: Guiding a picture with a pose, depth or edges; preprocessors.
- Reference images: IP-Adapter, Redux and editing models.
- Upscaling: Making a picture larger.
Model families
- Model families: SD 1.5, SDXL, Pony, Flux, Wan and the rest; what fits with what.
- Why families do not mix: What happens when files of different families meet.
- SD 1.5: The classic 512-pixel model.
- SDXL: The 1024-pixel model with two text encoders.
- Pony: An SDXL line with score tags.
- Illustrious: An anime SDXL line.
- SD 3 and 3.5: Three text encoders, 16-channel latent.
- Flux.1: dev, schnell and the tool models.
- Krea 2: Krea AI's own picture model; not FLUX.1 Krea [dev].
- Flux.2: The second Flux generation, dev and Klein.
- Chroma: An open model made from Flux.1 schnell.
- Qwen-Image: Text in pictures, and editing.
- Z-Image: A fast 6B image model.
- Kolors: An SDXL-like model with a ChatGLM encoder.
- AuraFlow: An open flow model.
- Stable Cascade: A model that works in stages.
- Wan 2.1: Open video, 1.3B and 14B.
- Wan 2.2: The video model of the lessons.
- HunyuanVideo: Tencent's video model.
- LTX-2: Video with sound.
Video and sound
- Video models: How video generation differs from pictures.
- Video length and frame rate: Frames, length and fps per model.
- Wan 2.2 two passes: Two experts, two samplers, one generation.
- Speed-up LoRAs: Four steps instead of twenty.
- First and last frame: A clip between two pictures.
- Video to video: Motion, depth or camera from another video.
- Pose animation: A character moves like a person in a video.
- Talking portraits: A face speaks a recording, with consent.
- Audio: Sound in ComfyUI.
Practice
- Graphics card memory: Why memory runs out and what helps.
- Node packs: Custom nodes, installing them with ComfyUI Manager, and the risk.
- Unsafe model formats: Why .safetensors, and why old formats can run code.
- Paid API nodes: Partner Nodes that send data out and cost credits.
- Notes, reroutes and primitives: Helper nodes and groups.
- Bypass and mute: Switching a node off without deleting it.
- Subgraphs: Several nodes packed into one.
- Reading a workflow: A method for someone else's graph.