# ComfyUI tutorial: install Comfy Desktop and make your first image > ComfyUI is a free, open source app that makes AI images on your own computer by joining boxes called nodes. Install Comfy Desktop from [comfy.org/download](https://www.comfy.org/download) and choose Local at first launch. The default text-to-image workflow opens. Download the model it asks for, type what you want in the positive prompt and press Run. Mindket editors · Published 2026-11-05 · Updated 2026-11-05 · https://mindket.com/learn/ai-images-video/comfyui-tutorial ## In short - Comfy Desktop installs ComfyUI, Python and everything they need, so there is no command line. - Models are separate downloads of several gigabytes, kept in ComfyUI/models. - Press R after adding a model so the loader node lists it. - Every image ComfyUI saves carries its workflow, so dragging it back in rebuilds the whole graph. ## Before you start: what ComfyUI is and what you need The tools in the earlier lessons run on Google's or OpenAI's servers. ComfyUI runs on your own computer. You pick an open model, download it once, and every picture after that is made on your machine, offline if you like. Instead of one prompt box, ComfyUI shows a canvas of boxes called nodes, joined by lines called links. One node loads the model, one holds your prompt, one makes the picture, one saves it. That looks busy at first. It is also why you can see, and change, every step. You install an app called [Comfy Desktop](https://docs.comfy.org/installation/desktop/overview). It installs ComfyUI for you with its own Python setup, and it can keep several copies side by side, each with its own version and add-ons. Its usage data, which you can turn off, never includes your workflows, prompts or pictures, according to Comfy's docs. Check your computer against the table first. A graphics card makes a big difference in speed. Without a strong one, the setup screen may badge Comfy Cloud, the hosted version, as recommended for your hardware. | System | Version | Hardware | | --- | --- | --- | | Windows | Windows 10 or later, x64 or ARM64 | A dedicated NVIDIA or AMD graphics card is recommended, not required | | macOS | macOS 13 Ventura or later | Apple Silicon, M1 or later | | Linux | Official .deb or AppImage package | See Comfy's Linux page | | Every system | At least 4.85 GB of disk per installation | 8 GB of memory, 16 GB recommended, plus room for models | ## Install Comfy Desktop, step by step Plan for a download and an install of a few gigabytes, then another few gigabytes for the first model. A wired connection or a quiet evening helps. At first launch you get a full-screen setup that collects every choice before anything is installed. If you close it halfway, the same screen comes back next time, so nothing is half done. The [Windows](https://docs.comfy.org/installation/desktop/windows) and [macOS](https://docs.comfy.org/installation/desktop/macos) pages show each screen. On Linux, install the official .deb package or the AppImage instead. Updates take care of themselves. When a new version is ready, a Desktop Update Ready button appears at the top of the window, and clicking it restarts the app and installs the update. ComfyUI itself, the engine inside the app, changes quickly. Its last three releases on [GitHub](https://github.com/Comfy-Org/ComfyUI/releases) came out eight days apart, and the project page there shows how widely it is used. **ComfyUI on GitHub, 5 November 2026** - **145.2k** stars from GitHub users (Source: [GitHub: Comfy-Org/ComfyUI](https://github.com/Comfy-Org/ComfyUI), read 2026-11-05) - **17.3k** forks, copies people made to change the code (Source: [GitHub: Comfy-Org/ComfyUI](https://github.com/Comfy-Org/ComfyUI), read 2026-11-05) - **0.44.0** latest release, out 4 November 2026 (Source: [GitHub: ComfyUI releases](https://github.com/Comfy-Org/ComfyUI/releases), read 2026-11-05) 1. Go to [comfy.org/download](https://www.comfy.org/download). The page detects your system and offers the matching installer. 2. On Windows, run the downloaded .exe installer, follow the prompts and open Comfy Desktop from the Start menu. On a Mac, open the .dmg file, drag Comfy Desktop into Applications and launch it. 3. If macOS shows a security warning, open **System Settings**, then **Privacy & Security**, and click **Open Anyway**. 4. At **How do you want to run Comfy?**, choose **Local** to install on this computer. Cloud runs ComfyUI on Comfy's servers instead. 5. Tick **I agree to the EULA and Terms of Service**. The two boxes below it, for anonymous usage data and beta features, are optional and ticked by default, so clear them if you prefer. 6. Leave **Express Install with recommended settings** cleared the first time, and click **Continue**. The install wizard then shows the ComfyUI version, the install folder and the startup options before anything is written to disk. 7. Accept or change those settings and wait for the install to finish. Then click the new instance card on the dashboard to launch ComfyUI. > When the app runs in Chinese, setup adds one more question, Enable Chinese Mirrors? Choosing Enable Mirrors fetches code and Python packages from mirrors that are faster in mainland China. ## Make your first image ComfyUI usually opens with its default text-to-image workflow already on the canvas. If the canvas is empty, click the **Templates** icon in the sidebar, or choose **Workflow**, then **Browse Workflow Templates**, and pick a text-to-image template. [Templates](https://docs.comfy.org/interface/features/template) are the quickest start, because each one knows which model files it needs. Most installs come with no model at all. When one is missing, ComfyUI marks the loader node in red and shows a small banner that counts the errors, with a View details link. The default workflow asks for an SD1.5 model, [v1-5-pruned-emaonly-fp16.safetensors](https://docs.comfy.org/get_started/first_generation), which Comfy's text-to-image tutorial puts at about 4GB. The sample below is Mindket's, written in the comma-and-phrase style SD1.5 reads best. Paste the first part into the positive prompt box and the second into the negative one. ````text Positive: sourdough loaf on a wooden counter, morning window light, rising steam, close-up, (golden hour:1.2), shallow depth of field, high detail Negative: low quality, blurry, deformed, text, watermark ```` - **Fix the missing model.** Click **View details** on the banner. In the Errors tab, click **Download** next to the model, or **Download all**. Comfy Desktop saves it into ComfyUI/models/checkpoints. - **If the download stalls,** copy the model link from the error details, download the file in your browser and move it into that checkpoints folder yourself. - **Refresh the lists.** Press **R** so the **Load Checkpoint** node sees the new file, then pick v1-5-pruned-emaonly-fp16 in it. The box must not read null. - **Write the prompt** in the two **CLIP Text Encode** boxes. The one wired to positive holds what you want, and the one wired to negative holds what you do not. - **Run it.** Click **Run** or press **Ctrl + Enter**. Each node lights up in turn while it works. - **Keep the picture.** It appears in the **Save Image** node, where a right-click saves it. A copy also goes to the ComfyUI/output folder. > Menus come in English, Chinese, Russian, French, Japanese and Korean. To switch, open Settings, then Comfy, then Locale. ## How nodes and links work A node is a box that takes something in, does one job and passes something out. Links carry the result from one node to the next. You never write code, because each node holds the code for its job, as Comfy's [nodes page](https://docs.comfy.org/development/core-concepts/nodes) puts it. Colors tell you what can connect. A link only joins an output to an input of the same color, which is the same data type: yellow for the CLIP model, pink for a latent image, blue for a finished pixel image. Drag from an output dot to a matching input dot to connect them. Nodes also show their state. A running node is highlighted while it works. A node with a bad input turns red after a run. A node that a downloaded workflow needs but your copy lacks shows as missing, and the fix is usually an update or a custom node. Right-click a node to change its mode. Never stops it from running, so the nodes after it get nothing and show errors. Bypass also skips it but lets the data flow past untouched, which is a quick way to switch a LoRA off without deleting it. The default text-to-image workflow has six nodes, and the table follows the picture through them from left to right. | Node | What it does | | --- | --- | | Load Checkpoint | Loads the model file, which holds three parts: the image model, a CLIP text encoder and a VAE | | Empty Latent Image | Sets the canvas size, so it decides the width and height of the picture | | CLIP Text Encode | Turns your positive and negative prompts into a form the model understands | | KSampler | The core of the workflow: it starts from random noise and removes it step by step, guided by the prompts | | VAE Decode | Turns the finished latent image into a normal picture | | Save Image | Shows the picture and saves it to the ComfyUI/output folder | > Every image ComfyUI saves holds its workflow inside the file. Drag one onto the canvas, or open it with Workflows and then Open, and the graph comes back as it was. ## Change the picture: prompts and KSampler settings Prompting here feels different from Gemini or ChatGPT. SD1.5 is an older, smaller model, and Comfy's [text-to-image tutorial](https://docs.comfy.org/tutorials/basic/text-to-image) gives it its own rules: write in English, use short phrases split by commas, and be specific. Weight a phrase by wrapping it, as in (golden hour:1.2), to make it more likely to show. The negative box is where unwanted things go. Write them as a plain list, such as blurry, extra fingers, watermark. Most of the other controls live in the KSampler node. Change one at a time and run again, so you can see what each one does. - **seed:** the starting noise. ComfyUI picks a new random seed after each run by default, which is why two runs give two pictures. Fix the seed to compare other settings fairly. - **steps:** how many rounds of noise removal. More steps add finer detail and take longer. - **cfg:** how strictly the picture follows the prompt. Too high a value overdoes it. - **sampler_name and scheduler:** the method and pace of the noise removal. Leave the defaults until the rest feels familiar. - **denoise:** how much noise to start from. At 1.0 the picture starts from pure noise, and lower values keep more of an input image, which matters once you edit pictures in ComfyUI. - **Canvas size:** set it in Empty Latent Image. SD1.5 was trained on 512×512 pictures, and quality drops when you ask it for a large square straight away. > Hands and complicated lighting are on Comfy's own list of SD1.5's weak spots. Try a different seed before you rewrite the prompt. ## Add more models Models are the files that do the work. The [models page](https://docs.comfy.org/development/core-concepts/models) names five common kinds: checkpoints, VAEs, LoRAs, ControlNets and upscalers. None come with the app, and a single model is often many gigabytes. To add one, download it from a site such as [Hugging Face](https://huggingface.co/models), [Civitai](https://civitai.com) or a project's GitHub page, and put it in the matching folder under ComfyUI/models, such as checkpoints, loras or vae. Then pick it in the loader node for that type. Loader nodes usually have names that start with Load. The Model Library in the left sidebar lists every model ComfyUI has found, and Comfy Desktop shares models across all your installs by default, so one download serves them all. The template library is the safer route for newer models. When ComfyUI adds first-class support for a model, a template usually appears with the right nodes already joined, and its error panel downloads the files into the right folders. Two cautions before you go further. Many models come through custom nodes, which are add-ons written by other people, so install only ones you trust and follow each project's instructions. And read the license on each model's download page before you use its pictures for work. - **LoRA:** a small add-on that pushes a checkpoint toward a style, a character or a concept. - **ControlNet:** extra guidance from an edge map, a depth map or a pose. - **Inpainting model:** fills or replaces an area inside an existing picture, like the edits in lesson 2. - **Removing a model:** there is no uninstall button yet. Delete its file from the models folder. ## What fofr and the ComfyUI team suggest after your first image fofr, who works on AI models at Google DeepMind by his X bio, posted a tip for beginners [on X](https://x.com/fofrAI/status/1813158025085046911) on July 16, 2024. His advice is to clear the canvas and join the nodes yourself, one small piece at a time, instead of only running finished graphs. The post is more than two years old and still holds: Comfy's [shortcuts page](https://docs.comfy.org/interface/shortcuts) lists Backspace as Clear workflow, and a double-click on the empty canvas opens a search for nodes to add. Try it once the default workflow has made a picture. That picture carries the whole graph, so dragging it back onto the canvas restores everything if you get lost. Clear the canvas, then add the six nodes from the table above one by one, joining each output to the input of the same color. When Run works, you understand every box in the default workflow. The ComfyUI team's own account [on X](https://x.com/ComfyUI/status/2001742683916955939) introduced a new template library on December 18, 2025, with workflows built for creative ideas and real tasks rather than only for testing a model. So when you want a result more than a lesson, open the template browser first and pick the closest workflow. It checks which model files it needs and offers to download them. In the video below, Max Novak sets ComfyUI up on his own PC and works through templates such as FLUX.2 for text-to-image and image-to-image. > Start from an empty workflow > > fofr, works on AI models at Google DeepMind, [X](https://x.com/fofrAI/status/1813158025085046911) (read 2026-11-05) [![fofr (@fofrAI) on X, 16 July 2024: ComfyUI beginners tip. Start from an empty workflow. Learn to wire up flows from scratch. Start small, build up.](https://mindket.com/blog/shots/x-fofrai-comfyui-tip.webp?v=23d41e4e475b)](https://x.com/fofrAI/status/1813158025085046911) fofr on X, 16 July 2024 [![ComfyUI (@ComfyUI) on X, 18 December 2025: Meet the new Template Library in ComfyUI. We now have workflows designed for creative ideas and real tasks, not just model experiments. Build faster and stay in control.](https://mindket.com/blog/shots/x-comfyui-template-library.webp?v=516d13602ce9)](https://x.com/ComfyUI/status/2001742683916955939) ComfyUI on X, 18 December 2025 [Get Started in ComfyUI w/ Max Novak: Beginner Tutorial/Guide](https://www.youtube.com/watch?v=tuXveeHJpDA) NVIDIA Studio on YouTube, February 2026: digital artist Max Novak installs ComfyUI, runs it locally on an NVIDIA GeForce RTX GPU and finds his way around the node canvas. > Screenshots of X's own embeds of the posts, taken on 5 November 2026, with the time in UTC. Select one to open the post on X. ## Common mistakes - Pressing Run before the model is in place. Clear the red loader node and the error banner first. - Saving a model in the wrong folder. A checkpoint goes in checkpoints and a LoRA in loras, or the loader will not list it. - Forgetting to press R after adding a file while ComfyUI is open. - Writing a long sentence prompt for SD1.5. Short phrases split by commas suit it better. - Changing four settings at once. When the picture improves, you cannot tell which change did it. - Asking SD1.5 for a large picture straight away. Start near its training size and upscale later. - Closing the window while a picture is still being made. Closing an instance window stops that copy of ComfyUI, and closing the last window quits Comfy Desktop. - Installing every custom node a downloaded workflow asks for without checking who wrote it. ## Next course: Design with AI This is the last lesson of AI images and video. You can now write prompts, edit photos, make short videos and run a model on your own computer. To share a workflow, choose Workflows and then Export, which saves it as a JSON file. Anyone with ComfyUI can open it, and the template library is the place to look for the next model to try. If you want designs rather than single pictures, the next course, [Design with AI](https://mindket.com/learn/ai-design/ai-logo-design), makes logos, social graphics, readable text in images and product mockups. ## Check yourself 1. You added a model file while ComfyUI was open. What makes it show up in the Load Checkpoint list? A. Press R to refresh B. Reinstall Comfy Desktop C. Rename the file to model.ckpt Answer: A. Pressing R refreshes the model lists, so the Load Checkpoint node sees the new file. 2. Which node turns your positive and negative prompts into a form the model can use? A. Save Image B. Empty Latent Image C. CLIP Text Encode D. VAE Decode Answer: C. CLIP Text Encode turns your prompts into a form the model understands. 3. You run the same workflow twice and get two different pictures. Why? A. The model file is damaged B. KSampler used a different random seed C. ComfyUI ignored the prompt Answer: B. By default the seed changes after each run, and a new seed means new starting noise. 4. Where should a downloaded LoRA file go? A. ComfyUI/output B. The desktop C. The folder that holds the installer D. ComfyUI/models/loras Answer: D. Each model type has its own subfolder under ComfyUI/models, and LoRAs go in loras. ## Frequently asked questions ### Is ComfyUI free? The software is. Comfy Desktop is open source, with its code on GitHub under the AGPL-3.0-or-later license or a commercial license from Comfy Org. Running models locally uses your own hardware. Comfy Cloud, the hosted option, has its own plans and terms. ### Do I need an NVIDIA graphics card for ComfyUI? No. On Windows, a dedicated NVIDIA or AMD card is recommended for good speed but not required, and Macs need Apple Silicon. Comfy's tutorial says SD1.5 runs smoothly on consumer cards with about 6GB of video memory. ### How do I uninstall ComfyUI completely? On Windows, remove Comfy Desktop in Settings, Apps, Installed apps. That removes only the launcher. Your installs, shared models, pictures and settings stay in folders under AppData, and you delete those by hand once you are sure you no longer need them. ### Can ComfyUI make videos too? Yes, with a video model. The template library lists the models ComfyUI supports, and Comfy's agent tools pages describe making video, audio and 3D as well as images. Expect much larger downloads than SD1.5. ## Sources - [Comfy Docs: Comfy Desktop overview](https://docs.comfy.org/installation/desktop/overview) (read 2026-11-05) - [Comfy Docs: Comfy Desktop for Windows](https://docs.comfy.org/installation/desktop/windows) (read 2026-11-05) - [Comfy Docs: Comfy Desktop for macOS](https://docs.comfy.org/installation/desktop/macos) (read 2026-11-05) - [Comfy Docs: Getting started with AI image generation](https://docs.comfy.org/get_started/first_generation) (read 2026-11-05) - [Comfy Docs: Workflow templates](https://docs.comfy.org/interface/features/template) (read 2026-11-05) - [Comfy Docs: Nodes](https://docs.comfy.org/development/core-concepts/nodes) (read 2026-11-05) - [Comfy Docs: Models](https://docs.comfy.org/development/core-concepts/models) (read 2026-11-05) - [Comfy Docs: Text to image workflow](https://docs.comfy.org/tutorials/basic/text-to-image) (read 2026-11-05) - [Comfy Docs: Keyboard shortcuts](https://docs.comfy.org/interface/shortcuts) (read 2026-11-05) - [GitHub: Comfy-Org/ComfyUI](https://github.com/Comfy-Org/ComfyUI) (read 2026-11-05) - [GitHub: ComfyUI releases](https://github.com/Comfy-Org/ComfyUI/releases) (read 2026-11-05) - [fofr on X: ComfyUI beginners tip](https://x.com/fofrAI/status/1813158025085046911) (read 2026-11-05) - [ComfyUI on X: the new Template Library](https://x.com/ComfyUI/status/2001742683916955939) (read 2026-11-05) - [NVIDIA Studio on YouTube: Get Started in ComfyUI w/ Max Novak](https://www.youtube.com/watch?v=tuXveeHJpDA) (read 2026-11-05) [Browse MCP servers](https://mindket.com/mcp-servers)