ComfyUI for Beginners: Install, Run Locally and Make Images, Video and Music
%20(2).avif)
%20transparent.avif)
ComfyUI for Beginners: Install, Run Locally and Make Images, Video and Music
ComfyUI is the free engine behind most open-source AI image and video tools. It is powerful, but when you first open it, it looks like a wall of boxes and wires. This guide walks you through it from zero: what it is, how to run it on your own PC, and how to make an image, a video and a song. Every time quoted here was measured on one Windows laptop with two graphics cards, and each one is a single recorded run, so treat the numbers as a starting point for your own tests.
What is ComfyUI? Three words that make it easy
Three words explain almost everything: model, workflow and interface.
- A model is the trained file that knows how to make pictures, video or sound.
- A workflow is the recipe that uses it. It loads the model, reads your prompt, generates, and saves the result.
- The interface is how you control that recipe.
In ComfyUI the recipe is drawn as boxes called nodes, joined by wires called links. One node loads the model, one turns your words into something the model understands, the sampler does the generating, one turns the result into an image, and one saves it.
Promptus gives you three ways into the same recipe: Playground (a simple form), CosyTemplates (ready-made recipes you can install with one click) and the Comfy UI tab (the full canvas).
Run it locally or online
In Promptus, a switch at the top right chooses where a job runs. Local runs it on your own graphics card. Online runs the same job on Promptus's servers, so you don't need a powerful card. When you open Model and Options, a model with a LOCAL tag, such as Z-Image Turbo, confirms the job is running on your PC. When Local is on, Promptus sends your prompt to ComfyUI on your own machine (localhost, port 8288) and saves the finished file in a folder called Comfy Output.
How to install ComfyUI with Promptus Manager
Promptus Manager is the control room for everything that runs on your PC. Open it from your Profile page. Its main window installs and starts the pieces: the GPU server, the CPU server, ComfyUI, and the Cosyflow server that runs templates. Green means running. Open the menu and choose ComfyUI to see the engine's status and live log. If a service won't start, read that log first: it shows the difference between a process that has started and a server you can actually use.
Make your first AI image
With the engine running, type a simple prompt, make sure you're on Local, and press Generate. In our test the prompt was "a young man with short dark hair, smiling in a sunlit café, thirty-five millimetre photo". On the RTX 5090, Z-Image Turbo took about 15 seconds from a cold start and about 7 seconds once the model was loaded. The first run is slower because the model has to load into graphics memory; after that it is already there.
Three tools make prompts easier: Prompt History (every prompt you have used, saved with its date), Enhance Prompt (adds detail such as light, background and camera angle, so read what it adds) and Describe Image (writes a prompt from a picture you already have).
Install ComfyUI workflows in one click
In plain ComfyUI, a new workflow means downloading a file, hunting for the right models and installing extra nodes. In Promptus, workflows come packaged as CosyTemplates. At the time of recording there were 256 of them, and 23 were installed on the test PC. Filters show where each one can run: Available Online, Available Locally, or Can Install Locally. Choose Install Locally on a card and Promptus downloads the template and its model; the small Fast GAN Video Upscaler took about 43 seconds.
ComfyUI nodes and the canvas, explained
Open the Comfy UI tab and you see the same Z-Image workflow you just used from the form. Follow the links from left to right: the model loads, your prompt is encoded, the sampler generates, the decoder turns the result into an image, and Save Image writes it to disk. Nodes can be moved without changing anything. To add one, open the Node Library, drag in Preview Image, and connect the decoder's IMAGE output to its images input. Outputs are on the right and inputs are on the left, which is the whole grammar of ComfyUI.
The Promptus part is the Cosyflow Editor. It decides how a workflow appears as a simple form. Each field gets a visibility level from 1 (Hidden) to 5 (Always), so you can hide the complicated settings and let anyone use your workflow without seeing a single node.
Make AI video and music locally
For video, open LTX-2.5 Image to Video from Available Locally and run it in Playground. Upload your first frame, then describe the motion rather than the picture: wind moves her hair, she turns toward the camera, slow push-in, warm golden light. On the RTX 5090, a five-second clip took 81 seconds. Check the model name and the LOCAL tag before you press Generate, because the Video tab can start on an online model.
For music, open the Audio tab, choose ACE-Step 1.5 under Local Models, write lyrics and add tags that describe the sound. About a minute later, a two-minute song was in the gallery.
Keep the same character across images
Mom is a chat for building a character. You describe who she is and what must stay the same, then change the scene and keep the same face. Its Iterate buttons use online models such as Nano Banana or GPT Image 2, and your chats are saved so you can return to a character any time.
Save and publish your AI images
Comfy Output holds everything ComfyUI has saved on your PC; My Collection holds the work you choose to keep. Add a result to your collection and it arrives marked Private. Choose Make the image public and it appears on your public gallery page. If Comfy Output looks out of date, Rebuild Database rescans the folder; it never deletes your pictures.
ComfyUI low VRAM: 8 GB vs 24 GB graphics cards
VRAM is your graphics card's own memory. Models live there while they work, so more VRAM means faster and bigger jobs. We ran the same jobs on both cards in the test laptop:
| Job | RTX 4060 (8 GB) | RTX 5090 (24 GB) |
|---|---|---|
| Z-Image picture | 33 seconds | 15 seconds (cold start) |
| 5-second LTX video | about 8 minutes (fast mode), about 4 minutes (low VRAM mode) | 81 seconds |
On the 8 GB card the video job peaked at about 7.9 GB in use. So an 8 GB card can make video; it is just slower, and low VRAM mode nearly halved the wait.
Does Sage Attention make ComfyUI faster?
Sage Attention replaces one heavy step in the model with a faster version. We tested it with the same prompt, seed and settings, with models unloaded first. On this laptop, from a cold start, it made both jobs slower: the picture took 15 seconds without it and 38 with it, and the video took 81 seconds without it and 92 with it. A speed setting that helps one machine can hurt another, so time your own workflow, change one thing, and time it again.
How to fix missing models in ComfyUI
If a box turns red when you open a workflow, a model file is missing. In Promptus, click View details: the Issues panel names the exact file, what it is for and its size, with a Download button (in our example, Download all was 4.84 GB). To see what is already on your drive, open Promptus Manager, then Models, where every installed model is listed with its size and you can delete what you don't need. Video models are large; the Wan 2.2 video files in the test were over 13 GB each.
How to automate ComfyUI with the API
Everything you do by clicking can also be done by a program. Promptus runs two local addresses on your PC: ComfyUI at localhost port 8288, and the Cosy server, which runs templates, at port 8190. The pattern is simple: send the workflow, get back a job number, ask for the result, and download the file. That makes batches possible (send fifty prompts instead of typing them) and lets an AI agent drive the app for you. In our test, four portraits sent one after another took about 30 seconds for the first, because the model had to load, and about 7 seconds each for the rest.
Frequently asked questions
Do I need a powerful graphics card to use ComfyUI locally?
A bigger card is faster, but in our test an 8 GB RTX 4060 produced both an image (33 seconds) and a five-second video (about 4 minutes in low VRAM mode).
Does ComfyUI run on Mac?
The video was recorded on Windows, and Promptus also runs on Mac.
Is Sage Attention worth turning on?
Not automatically. On the test laptop it was slower from a cold start, so measure it on your own machine.
Summary
Local keeps the work on your own PC. Templates give you ready-made recipes. The canvas shows you how they work. Manager shows you what is running. And your own timings tell you what to change. Watch the full video for every step on screen.
