Generate images with Stable Diffusion and Flux using ComfyUI, on a PC or rented GPU

By Mouhamed BANKOLE, CEO at ITNET Technologies · Updated November 6, 2026 · Reading time: 9 min

Generating visuals with Stable Diffusion or Flux on your own machine, with no credits to buy and no image sent to an online service: that is what ComfyUI offers. Its node-based interface is puzzling at first, then becomes the most flexible tool for chaining model, prompt, settings and edits. The real obstacle is the graphics card. Here is how ComfyUI works, how to install it, how much video memory you need, how to set it up on a rented GPU, including the compatibility pitfall on V100 cards, and how to reach it without exposing it.

In short

• ComfyUI is a free, open-source interface and engine (GPL-3.0 license) for generating images with diffusion models such as SD 1.5, SDXL, SD3.5 or Flux.
• You build a "workflow" by linking nodes: model loading, prompt, sampler, image saving.
• VRAM makes the difference: ComfyUI can run large models with little memory thanks to offloading, but much more slowly.
• On an NVIDIA V100, install PyTorch built with CUDA 12: CUDA 13 tools no longer target the Volta architecture.
• Do not expose ComfyUI to the Internet: reach it through an SSH tunnel.

Contents

1. What is ComfyUI?
2. How does a ComfyUI workflow work?
3. Which image models does ComfyUI support?
4. How do you install ComfyUI?
5. Which graphics card do you need for ComfyUI?
6. How do you install ComfyUI on a rented GPU?
7. How do you access ComfyUI remotely and safely?
8. ComfyUI or another interface: which should you choose?
9. How much does image generation cost on a rented GPU?

What is ComfyUI?

ComfyUI is a graphical interface, an API and an engine for diffusion models, organized as a graph of nodes. The project, published under the GPL-3.0 license on GitHub, describes itself as the most modular interface of its kind.

What you can do with it:

• generate images from text (text-to-image);
• transform or retouch an existing image (image-to-image, inpainting);
• chain several steps: generation, upscaling, correction;
• automate through its API to produce series of visuals.

ComfyUI belongs to the same family of local AI tools as Ollama, which runs text models.

How does a ComfyUI workflow work?

A ComfyUI workflow is a diagram where each node performs one step (load the model, encode the prompt, generate, decode, save) and links pass the result on to the next node. You see the whole process at a glance and can change a single step.

The basic workflow includes:

1. a model loader node (checkpoint);
2. two prompt nodes: what you want to see and what you want to avoid;
3. a sampler that generates the image from noise;
4. a decoder that turns the result into a visible image;
5. a save node that stores the file.

Pro tip: a workflow is saved inside the image file it produces. Drag a generated image into ComfyUI to recover the exact settings that created it.

Which image models does ComfyUI support?

ComfyUI supports the main families of diffusion models: Stable Diffusion 1.5, SDXL and SD3.5, Flux, Qwen Image, HiDream and others, as well as image editing models. The full list is in the official repository.

• Model files go in the models/checkpoints folder.
• Licenses vary from one model to another: check them before commercial use.

How do you install ComfyUI?

The simplest option is the ComfyUI Desktop app on Windows or macOS; on Linux or a server, you clone the repository, install PyTorch for your graphics card, then the dependencies. The following steps come from the official repository:

1. Clone the repository: git clone https://github.com/comfyanonymous/ComfyUI.
2. Create a dedicated Python environment.
3. Install PyTorch for your GPU (NVIDIA, AMD or Intel).
4. Install the dependencies: pip install -r requirements.txt.
5. Put a model in models/checkpoints.
6. Start ComfyUI: python main.py.

The Desktop app, available for Windows and macOS, avoids the command line entirely on a workstation.

Which graphics card do you need for ComfyUI?

A recent NVIDIA card with as much VRAM as possible is the most comfortable choice; ComfyUI can run large models with just 4 GB of VRAM and 8 GB of RAM, but by offloading part of the model, so slowly. That is what the ComfyUI repository states; it also supports AMD, Intel and Apple Silicon GPUs.

• Little VRAM: generation works but slowly, especially with Flux and high resolutions.
• Plenty of VRAM: heavy models fully loaded, faster batches of images.
• Quantized models reduce the memory needed, at the cost of a slight loss in quality.
• CPU and RAM matter less than the graphics card.

How do you install ComfyUI on a rented GPU?

Connect over SSH, check the card with nvidia-smi, install a PyTorch build compatible with your GPU, then ComfyUI and a model: on a V100, you must choose a PyTorch build made with CUDA 12. The CUDA 13.0 release notes from NVIDIA state that CUDA 13 tools can no longer target the Maxwell, Pascal and Volta architectures, while the 12.x series remains supported for them.

Procedure on a V100:

1. Check the GPU: nvidia-smi shows the card and driver.
2. Install PyTorch with CUDA 12, not the CUDA 13 variant the repository suggests by default.
3. Check compatibility: python -c "import torch; print(torch.cuda.get_arch_list())" should list sm_70.
4. Install ComfyUI and its dependencies as above.
5. Run it locally: python main.py, which only listens on the machine.

Wayhost GPU rental offers a dedicated NVIDIA V100 32 GB, hosted in Cergy, with root access by SSH key, at €0.60 excl. VAT per hour or €199.99 excl. VAT per month (checked on November 6, 2026). The 32 GB of VRAM let you load heavy models without offloading.

How do you access ComfyUI remotely and safely?

Keep ComfyUI listening locally on the server and open an SSH tunnel from your workstation: the interface then shows up in your browser as if it ran at home. ComfyUI has no user accounts: exposing it to the Internet means handing your GPU to anyone.

• Close the ComfyUI port in the firewall.
• Be careful with extensions ("custom nodes"): install them from trusted sources, because they run code on the server.
• Back up your workflows and images off the server.

ComfyUI or another interface: which should you choose?

ComfyUI suits you if you want to control every step and automate; a simpler interface is enough for an occasional image. The choice depends on your use, not on final quality, which depends mostly on the model.

• ComfyUI: reusable workflows, API, quick support for new models.
• Form-based interfaces: easier to learn, less flexible for chaining steps.
• Online services: nothing to install, but images and prompts go to a third party and credits must be paid.

For the text side of an AI project, our guide to hosting an AI application presents the possible architectures.

How much does image generation cost on a rented GPU?

The cost is GPU time: you pay by the hour during creative sessions, or by the month for continuous production, with no per-image billing. ComfyUI and most open models are free.

• Hourly: ideal for occasional sessions; shut the instance down when you are done.
• Monthly: for daily or automated generation.
• A whole machine: a physical server for a permanent workload.
• The interface alone: a VPS can host the application that sends requests to the GPU.

Our server hosting guide helps you choose between hourly rental, VPS and dedicated server.

Frequently asked questions

Is ComfyUI free?

Yes, ComfyUI is open-source software under the GPL-3.0 license. Image models have their own licenses, sometimes restrictive for commercial use. The only cost is the hardware or the GPU rental.

What is the difference between ComfyUI and Stable Diffusion?

Stable Diffusion is a model that generates images. ComfyUI is the interface and engine that run this model and others, such as Flux. So you use Stable Diffusion inside ComfyUI.

Can you use ComfyUI without a graphics card?

It is possible on the CPU, but generation becomes extremely slow. A dedicated graphics card is strongly recommended. Without a powerful GPU, renting one by the hour is often the most practical option.

How much VRAM does ComfyUI need?

ComfyUI says it can run large models with 4 GB of VRAM and 8 GB of RAM through offloading, but slowly. The more memory the card has, the faster heavy models such as Flux run. A 32 GB card loads models fully.

Does ComfyUI work on an NVIDIA V100?

Yes, provided you install PyTorch built with CUDA 12, because CUDA 13 tools no longer target the Volta architecture. Check that torch.cuda.get_arch_list() lists sm_70. The rest of the installation is the same.

Are ComfyUI extensions safe?

Not always: an extension runs code on your machine. Only install well-known extensions from their official repository. On a server, isolate ComfyUI and do not expose it to the Internet.

Where do models go in ComfyUI?

Main models go in ComfyUI's models/checkpoints folder. Other subfolders hold LoRAs, VAEs and upscaling models. A configuration file lets you share models with another interface.

Key takeaways

• ComfyUI runs Stable Diffusion, SDXL and Flux in a free node-based interface.
• A workflow links steps that you can change and reuse.
• VRAM decides comfort: little memory works, but slowly.
• On a V100, PyTorch must be built with CUDA 12, not CUDA 13.
• Access through an SSH tunnel only: ComfyUI has no user accounts.

To generate your images on an NVIDIA V100 32 GB hosted in France, discover Wayhost GPU rental.

Article written by Mouhamed BANKOLE, CEO at ITNET Technologies.