If you want to generate AI images on your own PC, two free apps come up again and again: ComfyUI, with its node-based workflow editor, and Stable Diffusion WebUI Forge, a faster-moving fork of the classic AUTOMATIC1111 web interface. Both run models such as Stable Diffusion and Flux locally on your graphics card, and both are open source. They're built for quite different users. Forge PC hasn't benchmarked the two apps against each other, so this comparison doesn't make speed claims; it's based on each project's own documentation.
The short answer
| ComfyUI | Forge |
|---|
Interface | Visual node graph, plus an App Mode for simplified front ends | Tabbed web UI (txt2img, img2img and so on), based on AUTOMATIC1111's SD WebUI |
Models (per each project) | Very broad: SD 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image and many more, plus video, audio and 3D models | Stable Diffusion checkpoints and Flux, including NF4 and GGUF quantised Flux models |
Hardware listed | NVIDIA, AMD, Intel, Apple Silicon and Ascend | One-click Windows package built on CUDA 12.1 and PyTorch 2.3.1 (NVIDIA) |
Installer | Comfy Desktop for Windows 10 and later and macOS, plus Windows portable and manual installs (including Linux) | One-click package with Git and Python included, or a manual install |
Development | Weekly release cycle (often changing around model launches) | Based on SD-WebUI 1.10.1; syncs with the original roughly every 90 days |
Best for | New models, video, complex or repeatable workflows, non-NVIDIA GPUs | AUTOMATIC1111 users and anyone who prefers forms to graphs |
ComfyUI: the most flexible option
ComfyUI describes itself as a modular diffusion model GUI, API and backend with a node graph interface. Each step of image generation, such as loading a model, encoding the prompt, sampling and decoding, is a node you connect. That looks intimidating at first, but it means you can see exactly what a workflow does, save it as a JSON file and reuse or share it. ComfyUI says images it generates can carry the complete workflow and seed, so you can reload a result's exact settings later.
Its strongest selling point is model support. The project lists a long range of image models, editing models such as Flux Kontext and Qwen Image Edit, video models including Wan 2.1 and 2.2 and LTX-Video, and audio and 3D models. New open models usually reach it quickly. Its developers also say it supports NVIDIA, AMD, Intel, Apple Silicon and Ascend hardware, and that it can run even the largest open models on as little as 4GB of VRAM plus 8GB of system RAM by streaming model weights. Treat that as the developers' claim: smaller graphics cards still mean slower generation and more compromises, and our low-VRAM ComfyUI guide covers what to change.
For newcomers, App Mode can turn a complex workflow into a simple front end, and the Comfy Desktop installer for Windows 10 and later handles set-up. Comfy's documentation recommends a dedicated NVIDIA or AMD graphics card for good performance (though it isn't required) and at least 4.85GB of disk space per installation, before models. Note that Comfy Desktop checks for updates when it starts and every 10 minutes, and these checks can't be turned off; you can only choose whether updates install automatically.
Our ComfyUI installation guide walks through set-up, and how to use LoRAs in ComfyUI is a good next step.
Forge: the familiar web UI, updated
Forge's developers describe it as a platform on top of Stable Diffusion WebUI (AUTOMATIC1111) to make development easier, optimise resource management, speed up inference and study experimental features. If you've used AUTOMATIC1111, it will feel immediately familiar: prompts, sliders and tabs, not nodes. Its README says it's currently based on SD-WebUI 1.10.1 and that, because the original WebUI is "almost static now", Forge syncs with it every 90 days or when there are important fixes.
The main addition over the original is Flux support. Forge says Flux BNB NF4 and GGUF Q8_0, Q5_0, Q5_1, Q4_0 and Q4_1 models are natively supported, with a GPU Weight slider and swap options to control how much of the model sits in VRAM. Its own troubleshooting advice is not to set GPU Weight too high.
The Windows one-click package includes Git and Python and is built on CUDA 12.1 and PyTorch 2.3.1, so it's aimed at NVIDIA graphics cards; the README doesn't offer an equivalent one-click package for AMD or Intel GPUs. If you install Forge as a branch of an existing AUTOMATIC1111 set-up, the README says you can reuse the checkpoints and extensions you already have. The Gradio 4 interface has one quirk worth knowing: you move the canvas with the right mouse button.
Which should you choose?
Choose ComfyUI if you want the newest image or video models, you have an AMD or Intel graphics card or a Mac, or you want repeatable workflows you can save and share. It has the steeper learning curve, but it's where new open models tend to appear first.
Choose Forge if you're coming from AUTOMATIC1111, you have an NVIDIA graphics card and you mainly want Stable Diffusion and Flux images through a simple form-based interface.
Neither choice locks you in: ComfyUI loads standard Stable Diffusion checkpoint files (ckpt/safetensors), and Forge's README says it can reuse the SD checkpoints from an AUTOMATIC1111 install, so you can try both with the same models. If you want something even simpler, see ComfyUI vs Amuse or our best local AI image generators. For hardware, our PC for Stable Diffusion and Flux guide explains how much VRAM to aim for.
Our verdict
ComfyUI is the better long-term choice for most people: it supports far more models, including video, runs on NVIDIA, AMD, Intel and Apple hardware, and its node workflows are saveable and repeatable. Forge is the easier switch for AUTOMATIC1111 users with NVIDIA cards who just want Stable Diffusion and Flux in a familiar interface. Forge PC hasn't benchmarked the two; this comparison is based on each project's documentation.