ComfyUI: The Node-Based Powerhouse for AI Image & Video Generation 🧩
If you've ever used AI image generation tools like Stable Diffusion or FLUX, you've probably seen two approaches: simple web UIs (like AUTOMATIC1111) and node-based systems where you connect boxes to build workflows. ComfyUI is the most popular node-based interface — and for good reason. It's powerful, modular, and supports almost every AI image and video model out there.
ComfyUI is a free, open-source tool (111k ⭐ on GitHub, 13k forks) that lets you build AI video generation and image pipelines by connecting ComfyUI nodes visually. Want to generate an image, upscale it 4x, apply a ControlNet, and then run it through a second model? You just drag, connect, and click generate. No coding required. It's like building with Lego blocks — but for AI.
In this ComfyUI tutorial, we'll cover the most popular ComfyUI models people run, from FLUX to SDXL to Wan Video. With active development (5,000+ commits, updated daily by creator comfyanonymous) and 3,600+ custom nodes available, ComfyUI supports everything from FLUX and SDXL to Wan 2.2 video and Hunyuan3D. Let's dive in. 🔍
🏆 Top Models Used with ComfyUI
🥇 FLUX.1-dev / FLUX.1-schnell — 12.7M ⭐ (dev) / 4.8M ⭐ (schnell) — 🖼️ Best overall image quality
Black Forest Labs' FLUX.1 is the gold standard for AI image generation in 2025. The dev variant (741K HF downloads) offers the best quality-to-speed ratio, while schnell (724K downloads) trades a bit of quality for 4-step inference speed. Both are first-class citizens in ComfyUI with dedicated nodes and optimizations. ComfyUI natively supports FLUX with dedicated sampling nodes, VAE handling, and advanced features like FLUX Fill for inpainting.
💾 VRAM: 6-8GB (schnell), 10-12GB (dev, Q4 quantized)
⚙️ Hardware: Consumer GPUs with 8GB+ VRAM (RTX 3070/4070). Schnell runs on old GPUs like GTX 1080 at reduced speed. Latest high-end GPUs (RTX 4090/5090) for dev at full precision.
🥈 Wan 2.2 Video — 5.9M downloads (Kijai repackaged) — 🎬 Best open video generation
Alibaba's Wan 2.2 is the most downloaded AI video model for ComfyUI — and by a wide margin. The ComfyUI repackaged version by Kijai has 6 million downloads, making it the most popular video model in the ComfyUI ecosystem. Supports text-to-video, image-to-video, and video-to-video workflows. Available in full precision, FP8 scaled, and even GGUF quantized versions for lower VRAM setups.
💾 VRAM: 8-12GB (FP8), 16-24GB (full precision)
⚙️ Hardware: Consumer GPUs with 12GB+ VRAM (RTX 3080/4080). Full precision needs latest high-end GPUs (RTX 4090/5090). GGUF versions can run on 8GB cards.
🥉 SDXL (Stable Diffusion XL) — 7.6K ⭐ on HF (base 1.0) — 🎨 The reliable workhorse
Stability AI's SDXL 1.0 is the most mature and battle-tested image generation model on ComfyUI. With 2 million downloads on HuggingFace, thousands of community fine-tunes, and years of refinement, SDXL remains the go-to model for reliable, high-quality generation. Supports everything — LoRAs, ControlNets, IP-Adapters, T2I-Adapters, and hundreds of community checkpoints like RealVisXL and Juggernaut XL.
💾 VRAM: ~6-8GB
⚙️ Hardware: Consumer GPUs with 6GB+ VRAM (RTX 3060/4060). Runs fine on old GPUs like GTX 1070/1080 with optimizations. The most accessible high-quality model for older hardware.
4️⃣ FLUX.2 Klein — 967K downloads — ⚡ Next-gen efficiency
The latest from Black Forest Labs — FLUX.2 Klein comes in 4B and 9B parameter variants. Klein (German for "small") is optimized for efficiency while maintaining FLUX-level quality. The klein-base-4B (967K downloads) is the most popular, offering a sweet spot between speed and output quality. Works natively in ComfyUI with custom sampling nodes.
💾 VRAM: ~4-6GB (4B), ~8-10GB (9B)
⚙️ Hardware: 4B variant runs on old consumer GPUs with 4GB+ VRAM (GTX 1060/1070). 9B variant needs consumer GPUs with 8GB+ VRAM.
5️⃣ SD3.5 Medium / Large — 274K / 52K downloads — 🔬 Stability's latest
Stability AI's Stable Diffusion 3.5 series brings a new architecture with better prompt adherence and text rendering. Available in Medium (2.5B params, 274K downloads) and Large (8B params, 52K downloads). ComfyUI has a dedicated FP8 optimized version (by Comfy-Org) that makes it runnable on consumer GPUs. Excellent for typography and complex prompts where SDXL struggles.
💾 VRAM: ~6-8GB (Medium FP8), ~12-16GB (Large)
⚙️ Hardware: Medium runs on consumer GPUs with 8GB+ VRAM. Large needs latest high-end GPUs with 16GB+ VRAM.
6️⃣ HunyuanVideo 1.5 — 401K downloads — 🎥 Tencent's video model
Tencent's HunyuanVideo 1.5 repackaged for ComfyUI (401K downloads) is a strong alternative to Wan for video generation. Supports text-to-video with good motion coherence and scene understanding. The ComfyUI version includes optimized workflows for lower VRAM usage. Updated regularly by the Comfy-Org team.
💾 VRAM: ~10-14GB
⚙️ Hardware: Requires latest high-end GPUs with 12GB+ VRAM (RTX 4090/5090). Not suitable for mid-range consumer GPUs.
7️⃣ FLUX.2-dev — 211K downloads — 🚀 Developer-grade FLUX
The developer-oriented FLUX model from Black Forest Labs. Offers the highest quality but requires more VRAM. Best for professional workflows where quality matters more than speed. ComfyUI supports it with optimized node sets for faster inference.
💾 VRAM: ~12-16GB (Q4), ~24GB+ (full precision)
⚙️ Hardware: Requires consumer GPUs with 12GB+ VRAM (RTX 4080). Full precision needs latest high-end GPUs (RTX 4090/5090/A100).
8️⃣ SDXL Turbo — 874K downloads — ⚡ Real-time generation
Stability AI's SDXL Turbo is distilled for 1-4 step inference — you get decent images in under a second. Perfect for real-time applications, rapid prototyping, and iterative workflows where speed is the priority. Works beautifully in ComfyUI with dedicated turbo sampling nodes.
💾 VRAM: ~4-6GB
⚙️ Hardware: Runs on old consumer GPUs with 4GB+ VRAM (GTX 1060/1070). Generates in ~0.5-2 seconds on modern GPUs.
9️⃣ Playground v2.5 — 239K downloads — 🎨 Aesthetic-focused model
Playground AI's v2.5 1024px-aesthetic model is fine-tuned specifically for visually appealing outputs — think Instagram-worthy images right out of the box. Requires less prompt engineering than SDXL for good results. Great for beginners who want quality without complex workflows.
💾 VRAM: ~6-8GB
⚙️ Hardware: Consumer GPUs with 6GB+ VRAM. Runs on old GPUs like GTX 1080 at reduced speed.
🔟 Stable Diffusion 1.5 — 1.4M downloads — 🕰️ The classic that started it all
The granddaddy of open-source AI image generation. SD 1.5 has the largest ecosystem of LoRAs, embeddings, hypernetworks, and fine-tunes of any model ever created — hundreds of thousands. While newer models produce better quality, SD 1.5 remains unbeatable for its speed, low VRAM requirements, and massive community support. Perfect for learning ComfyUI basics.
💾 VRAM: ~2-4GB
⚙️ Hardware: Runs on literally any modern CPU with integrated graphics (slow) or any old GPU with 2GB+ VRAM (GTX 1050/Tesla K80). The most accessible model ever made.
🎖️ Honourable Mentions
🧠 Qwen-Image — 1.4M downloads (ComfyUI repackaged)
Alibaba's Qwen-Image is a powerful vision-language model repackaged for ComfyUI workflows. Can understand and generate images based on complex natural language descriptions. Great for tasks that combine understanding and generation — like "make this image look like a painting from the Renaissance period."
💾 VRAM: ~6-8GB
⚙️ Hardware: Consumer GPUs with 8GB+ VRAM. Runs on latest high-end GPUs for faster inference.
🌊 Hunyuan3D 2.0 — ComfyUI supported
Tencent's 3D generation model that creates 3D models from single images or text descriptions. Available in full and low-VRAM versions on ComfyUI. Turns a photo of a chair into a full 3D mesh you can rotate and export.
💾 VRAM: ~6-10GB (low-VRAM version), ~16GB+ (full)
⚙️ Hardware: Consumer GPUs with 8GB+ VRAM for low-VRAM mode. Latest high-end GPUs for full quality.
🎵 Audio Models (AudioCraft / MMAudio)
ComfyUI has expanded beyond images and video — you can now generate music and sound effects using AudioCraft (Meta) and MMAudio nodes. Connect audio generation into your video workflows for complete AI content creation.
💾 VRAM: ~4-8GB
⚙️ Hardware: Consumer GPUs with 6GB+ VRAM. Smaller models work on modern CPUs.
💾 Hardware Requirements at a Glance
🖥️ Modern CPUs (no GPU needed)
ComfyUI can technically run on CPU but it's extremely slow — expect 2-10 minutes per SD1.5 image. Only viable for SD1.5 workflows. All modern models (SDXL, FLUX, Wan) require a GPU.
🕰️ Old Consumer GPUs (2-6GB VRAM)
Examples: GTX 1050/1060/1070/1080, GTX 1650/1660, AMD RX 480/580/590 (roughly any GPU from the GTX 1000 series onwards)
Can run: SD1.5, SDXL (with --lowvram flag), SDXL Turbo, FLUX.2 Klein 4B (FP8), FLUX.1-schnell (FP8), Playground v2.5 (low res)
Expect: 30-120 seconds per generation. Use --lowvram flag when launching ComfyUI. Avoid SD3.5 Large, Wan Video, and HunyuanVideo — they won't fit.
🎮 Consumer GPUs (8-16GB VRAM)
Examples: RTX 3060/3070/4060/4070/4080, AMD RX 7000 series
Can run: SDXL, FLUX.1-schnell, FLUX.2 Klein (all variants), SD3.5 Medium, Wan Video (FP8/GGUF), Qwen-Image, Hunyuan3D 2.0 (low-VRAM), HunyuanVideo (FP8), Audio models
The sweet spot for most users. Can run almost everything at reasonable speeds.
🚀 Latest High-End GPUs (24GB+ VRAM)
Examples: RTX 4090/5090, A6000, A100
Can run: FLUX.1-dev (full precision), FLUX.2-dev, Wan Video (full), SD3.5 Large, HunyuanVideo 1.5, all models at max resolution and quality
macOS users: Apple Silicon Macs (M1-M4) with 16GB+ unified memory can run ComfyUI with MPS acceleration. Supports SDXL, FLUX.1-schnell (MPS optimizations), and SD1.5 workflows. Video models (Wan, HunyuanVideo) are NVIDIA-only for now.
📦 How to Install ComfyUI
🪟 Windows
Download the ComfyUI standalone package from the GitHub releases page. Extract and run run_nvidia_gpu.bat (NVIDIA) or main.py directly. For AMD users, use the AMD portable script. The web UI opens at http://localhost:8188. To install custom nodes, use the built-in ComfyUI Manager (install via git clone into custom_nodes/).
🍎 macOS
Option 1 — One-click via Pinokio: Install Pinokio and search for "ComfyUI" — one click install.
Option 2 — Manual: git clone https://github.com/Comfy-Org/ComfyUI.git, then pip install -r requirements.txt. Launch with python main.py. Apple Silicon Macs get MPS acceleration automatically.
🐧 Linux
git clone https://github.com/Comfy-Org/ComfyUI.git && cd ComfyUI && pip install -r requirements.txt && python main.py
For NVIDIA GPUs, ensure CUDA toolkit is installed. For AMD GPUs, use the --rocm or PyTorch ROCm build.
🤖 Android
Not natively supported. Use a remote ComfyUI server on your desktop and access the web UI via your phone browser. Some lightweight workflows (SD1.5) can run via Termux on high-end Android devices but it's not practical.
🐳 Docker
Community Docker images available. The most popular setup:
docker run -d --gpus all -p 8188:8188 -v ./models:/workspace/ComfyUI/models comfy-docker/comfyui
Mount your models folder for persistence. Works on Linux natively, Windows via WSL2, macOS (CPU-only).
🎯 Essential Custom Nodes for ComfyUI
ComfyUI's real power comes from its custom node ecosystem. Here are the must-haves:
• ComfyUI Manager — install and update custom nodes from within the UI
• ComfyUI-Impact-Pack — advanced face detection, segmentation, and masking
• WAS Node Suite — text tools, image processing, video tools, and utilities
• Efficiency Nodes — simplified workflows for faster generation
• ControlNet Aux — preprocessors for ControlNet (depth, canny, openpose, etc.)
• AnimateDiff — generate animated GIFs and videos from Stable Diffusion
• IP-Adapter — image-prompting: "make an image in the style of this reference"
🔄 ComfyUI vs Alternatives
• AUTOMATIC1111 / SD WebUI Forge — simpler to use, better for beginners. Less flexible than ComfyUI for complex pipelines. Both can use the same models.
• ComfyUI — more powerful, node-based, supports more models. Steeper learning curve but infinitely flexible. Runs any model from SD1.5 to Wan Video.
• Pinokio — as covered in our earlier guide, Pinokio can install ComfyUI with one click. Think of Pinokio as the app store and ComfyUI as the app itself.
• InvokeAI — polished UI with canvas-based editing. Less model support than ComfyUI but better for iterative editing workflows.
🔮 Is ComfyUI Right for You?
If you're serious about AI image and video generation, ComfyUI is essential. Yes, the node-based interface looks intimidating at first — but once you understand the basics (model → VAE → sampler → latent → decoder → output), it clicks. And the flexibility is unmatched: you can do things in ComfyUI that no other tool can do, like chaining multiple models, building custom ControlNet pipelines, or generating video frame sequences from a single workflow.
For beginners, start with SDXL or FLUX.1-schnell and a simple 3-node workflow. Save community workflows from OpenArt or CivitAI and load them in ComfyUI to learn how they work. Within a few hours, you'll be building your own. 🚀
🤝 Alternatives to ComfyUI
• AUTOMATIC1111 Stable Diffusion WebUI — most popular traditional UI. Great for beginners.
• Stable Diffusion WebUI Forge — optimized fork of AUTOMATIC1111 with better memory management.
• InvokeAI — canvas-based editing with unified brush tools. Good for artists.
• Fooocus — minimalist UI focused on simplicity. No node-based workflows.
• DiffusionBee — macOS-only, very beginner friendly. Limited model support.
• Pinokio — app store that can install ComfyUI and other tools with one click.