A Dedicated GPU Server for ComfyUI, SDXL and FLUX
Stable Diffusion hosting here means a bare-metal server with its own NVIDIA card, on which you install ComfyUI and the models you choose. The RTX PRO 4000 server (24 GB, $449.99 a month) runs SDXL, Stable Diffusion 3.5 and FLUX.1; the RTX PRO 6000 server (96 GB, $1,656.15 a month) adds the video models that ask for 60 GB to 80 GB. Both are on the GPU dedicated servers page, with unlimited traffic and no per-image fee.
What is Stable Diffusion hosting on a dedicated server?
Stable Diffusion hosting: running an open-weight image model on a GPU you rent, instead of paying a service per image or per seat. On a dedicated server the whole card belongs to one customer, so generation speed does not depend on other tenants, the model files stay on your own NVMe drives, and prompts and outputs do not pass through a third-party image service.
The usual front end is ComfyUI, a node-based interface in which a workflow is a graph: load a checkpoint, encode the prompt, sample, decode, save. The same server can run more than one tool, because it is a plain Linux machine with root access. You choose the operating system at order (Ubuntu 24.04 LTS is the default; there is no Windows option), and it arrives clean: drivers, Python, ComfyUI and models are installed by you, with the steps further down.
What the server does not include: it is not a managed image service, there is no preloaded model library, and there is no hourly billing. If you want to pay only for the minutes you render, an hourly GPU cloud or a subscription is the better fit, and this page says when.
Which image and video models fit in 24 GB and in 96 GB?
File sizes and licences are from the model pages on Hugging Face, checked on 16 November 2026. The image model cards publish no vRAM figure, so the image rows compare file size with card memory; the video rows quote the memory the model authors state.
| Model | Size or stated memory | Licence | RTX PRO 4000, 24 GB, $449.99 | RTX PRO 6000, 96 GB, $1,656.15 |
|---|---|---|---|---|
| SDXL base 1.0 | 6.94 GB file | CreativeML Open RAIL++-M | Fits with room to spare | Fits, several pipelines at once |
| Stable Diffusion 3.5 Large (8B parameters) | 16.5 GB file, plus three text encoders | Stability Community License: free under $1M annual revenue, Enterprise licence above | File fits; the encoders, including T5-xxl, come on top, so test your workflow | Fits |
| FLUX.1 [schnell] (12B parameters, 1 to 4 steps) | 23.8 GB file, plus a 335 MB autoencoder | Apache-2.0 | The file alone is as large as the card: expect offloading or lower precision | Fits at full precision |
| FLUX.1 [dev] (12B parameters) | 23.8 GB file, plus a 335 MB autoencoder | FLUX.1 [dev] Non-Commercial License | As schnell | Fits at full precision |
| Wan 2.2 TI2V-5B (720P video, 24 FPS) | "at least 24GB VRAM" | Apache 2.0 | At the stated minimum | Fits |
| Wan 2.2 T2V-A14B (480P and 720P video) | "at least 80GB VRAM" | Apache 2.0 | Below the stated minimum | Fits |
| HunyuanVideo (13B+ parameters) | 45 GB peak at 544x960, 60 GB at 720x1280, 129 frames; 80 GB recommended | Tencent Hunyuan Community | Does not fit | Fits |
Licences matter more than memory for a business. FLUX.1 [schnell] and Wan 2.2 are Apache-licensed. FLUX.1 [dev] is released under a non-commercial model licence: its card says outputs may be used commercially, but the model licence itself is non-commercial, so read the licence text and speak to Black Forest Labs before you build a paid product on it. Stable Diffusion 3.5 Large is free for organisations under $1M in annual revenue and needs an Enterprise licence above that.
What 96 GB changes. Apart from the 60 GB to 80 GB video models, one RTX PRO 6000 can be divided with NVIDIA MIG into four isolated 24 GB instances, so four artists each get their own ComfyUI on one card. The RTX PRO 4000 does not support MIG. The Wan authors report a 5-second 720P clip from the 5B model in under 9 minutes on one consumer GPU; we publish no timing of our own, so run your workflow and time it. The RTX PRO 4000 vs RTX PRO 6000 comparison covers the two cards in detail.
Dedicated server, Midjourney or Runway seats, or hourly cloud: which costs less?
A server is a fixed cost and a subscription scales with people, so the comparison is a seat count. Midjourney's plans, checked on 16 November 2026, are Basic $10, Standard $30 (15 fast hours), Pro $60 (30 fast hours) and Mega $120 (60 fast hours) per month. Runway lists Standard at $15 (625 credits), Pro at $35 (2,250 credits) and Max at $95 (9,500 credits) per user per month billed monthly. Hourly cloud prices are on-demand rates of the same date.
| Option | Monthly cost | Equals the 24 GB server at | Notes |
|---|---|---|---|
| RTX PRO 4000 server, 24 GB | $449.99 | - | One-time setup $346.48, so the first month is $796.47. No cap on images. |
| Midjourney Standard | $30 per seat | 15 seats ($450) | 15 fast hours per seat |
| Midjourney Pro | $60 per seat | 8 seats ($480) | 30 fast hours per seat |
| Midjourney Mega | $120 per seat | 4 seats ($480) | 60 fast hours per seat |
| Runway Pro (as listed) | $35 per seat | 13 seats ($455) | 2,250 credits per seat |
| Runway Max (as listed) | $95 per seat | 5 seats ($475) | 9,500 credits per seat |
| AWS g5.xlarge, one A10G 24 GB, N. Virginia | $1.006 per hour | About 447 hours | 730 hours: $734.38. Storage and $0.09 per GB egress extra |
| AWS g6.xlarge, one L4 24 GB, N. Virginia | $0.8048 per hour | About 559 hours | 730 hours: $587.50. Storage and egress extra |
| Google Cloud g2-standard-4, one L4 24 GB, Iowa | $0.706832 per hour | About 637 hours | 730 hours: $515.99. Egress $0.12 per GiB on the Premium Tier |
Where a subscription fits. Below the seat counts in the table, paying per seat costs less, with nothing to install or patch and no setup fee. Midjourney's own model is not available as open weights, so if its look is what your clients want, a server does not replace it.
Where hourly cloud fits. If you render a few hours a day, well below the hours in the table, an instance you can stop is the better buy.
Where the server wins. Studios with 15 or more seats, pipelines that render all day (batch product shots, dataset generation, video), custom checkpoints and LoRAs that you train and keep, and work under NDA where prompts and reference images should stay on a machine only you use. Per hour over a 730-hour month the 24 GB server works out at $0.62 and the 96 GB server at $2.27.
Is it safe to expose ComfyUI to the internet?
Not directly. By default ComfyUI listens on 127.0.0.1, port 8188, which means only the server itself can reach it. Starting it with --listen and no argument makes it listen on all IPv4 and IPv6 addresses, and on a server with a public IPv4 address that puts the interface on the open internet.
A ComfyUI workflow can load files and run custom nodes, and custom nodes are Python code that runs with the rights of the ComfyUI process. Treat the port like a shell. Three ways to use it safely, in order of effort:
- SSH tunnel. Leave the default bind address and forward the port from your workstation:
ssh -L 8188:127.0.0.1:8188 user@your-server, then open http://localhost:8188. - Firewall allow-list. If a team needs access, bind to the public address but allow port 8188 only from your office or VPN addresses.
- Reverse proxy with TLS and a login. Put a web server with HTTPS and authentication in front and keep ComfyUI itself on localhost.
Run ComfyUI as an unprivileged user, not root, and install only custom nodes whose source you trust. The server is a single machine: keep copies of your workflows, LoRAs and outputs somewhere else, and consider software RAID 1 on the two NVMe drives.
Two servers for image and video generation
24 GB for image models at $449.99 a month, or 96 GB for video models and several users on one card at $1,656.15 a month.
NVIDIA RTX PRO 4000
- 770 AI TOPS
- CUDA and FP4 support
- 1 dedicated IPv4 included
NVIDIA RTX PRO 6000
- 3,511 AI TOPS
- CUDA, FP4 and Multi-Instance GPU (MIG) support
- 1 dedicated IPv4 included
Every RTX PRO server includes: a 1 Gbps port with unlimited traffic, one dedicated IPv4 address, hosting in an ISO 27001 certified data center, and your choice of Ubuntu, Debian, AlmaLinux, Rocky Linux, CentOS Stream, openSUSE or Arch Linux, or a rescue system to install your own OS. Billing runs monthly, quarterly, semi-annually, annually, biennially or triennially, and the setup fee is charged once with the first invoice.
How to install ComfyUI on the server
These steps assume Ubuntu 24.04 LTS, the default at order. The driver commands are NVIDIA's documented apt route and the ComfyUI commands are from its README, both read on 16 November 2026.
- Install the NVIDIA driver and reboot.
After the reboot,apt install linux-headers-$(uname -r) wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb dpkg -i cuda-keyring_1.1-1_all.deb apt update apt install nvidia-open rebootnvidia-smimust show the card and its 24 GB or 96 GB of memory. - Install ComfyUI as a normal user, in a virtual environment. The README requires a cu130 or newer PyTorch build on NVIDIA 20-series cards and later.
git clone https://github.com/comfyanonymous/ComfyUI cd ComfyUI python3 -m venv venv && source venv/bin/activate pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130 pip install -r requirements.txt - Add a model. Download a checkpoint such as
sd_xl_base_1.0.safetensors(6.94 GB) intomodels/checkpoints. A 23.8 GB FLUX file is about 190 gigabits, so it takes at least three minutes at the full 1 Gbps line rate. - Start it and connect.
python main.pystarts ComfyUI on 127.0.0.1:8188. Open an SSH tunnel from your computer and load http://localhost:8188. - Verify. Queue the default workflow and watch
nvidia-smiin a second terminal: memory use should rise on the card while the image samples.
The full walkthrough, including FLUX and a systemd unit, is in ComfyUI and FLUX on a GPU server; the older Stable Diffusion on a GPU server post covers the general setup.
What else can the same server do?
The card is not tied to one tool. The RTX PRO 4000 has two NVENC encoders and the RTX PRO 6000 four, with AV1 and 4:2:2 H.264 and HEVC encoding and no limit on concurrent sessions (NVIDIA support matrix, checked on 16 November 2026), so a render node can also transcode video with NVENC. Blender renders headless from the command line with Cycles on OptiX, as the Blender rendering guide shows. And a prompt-writing or captioning model can run beside the image model: a 24 GB card has room for SDXL and a small language model at once, and a 96 GB card for FLUX and a 70B model at 4-bit (43 GB). For language models as the main workload, see Ollama hosting.
Other GPU server pages
The same servers, sized for other workloads.
RTX PRO 4000 Server
The 24 GB card in detail: specifications, limits and what it runs for $449.99 a month.
96 GBRTX PRO 6000 Server
96 GB for video models, large language models and MIG.
LLMOllama Hosting
Language models next to your image pipeline, with one command to install.
384 GBMulti-GPU Server
Four 96 GB cards for a render or generation farm in one machine.
CostGPU Cloud Alternative
When a monthly server beats hourly GPU instances, and when it does not.
PrivacyPrivate AI Hosting
Keeping prompts, references and outputs on hardware one tenant controls.
Image and video generation guides
ComfyUI Server Setup: Run FLUX, SD 3.5 and Video Models on a Dedicated GPU Server
ComfyUI server setup on Ubuntu 24.04: the documented install, where the 23.8 GB FLUX.1 files go, what fits 24 GB and 96 GB, and safe remote access.
GuideRun Stable Diffusion on a GPU Server: Setup and Sizing
Run Stable Diffusion on a GPU server: what consumes VRAM, ComfyUI as a service, the memory-saving flags, matching card to workload and licence risk.
GuideThe First Hour on a GPU Server: NVIDIA Drivers, CUDA and Docker on Ubuntu 24.04
Install NVIDIA drivers on Ubuntu 24.04 server in 7 steps: nvidia-open from NVIDIA's repo, Docker, NVIDIA Container Toolkit 1.20.1 and a test container.
GuideRTX PRO 4000 vs RTX PRO 6000 Blackwell: 24 GB or 96 GB?
RTX PRO 4000 vs 6000 Blackwell: 24 GB at 70 W against 96 GB at 300 W, which models fit each, MIG on the 6000 only, and servers at $449.99 and $1,656.15 a month.
GuideNVENC FFmpeg Server: GPU Video Transcoding with H.264, HEVC and AV1 on a Dedicated Server
NVENC FFmpeg server setup: 4 encoders on the RTX PRO 6000 and 2 on the RTX PRO 4000 with no session cap, AV1 and 4:2:2, NVIDIA's commands and a batch script.
GuideBlender Render Server: Headless Cycles GPU Rendering with OptiX on a Linux Dedicated Server
Blender render server on Ubuntu 24.04: headless Cycles with OptiX from the command line, scenes on 24 GB vs 96 GB, rsync in and out, and server vs render farm.
Frequently Asked Questions
--listen without an argument opens it on every address, which you should not do on a public IP without a firewall rule.Run ComfyUI on your own GPU
24 GB from $449.99 a month or 96 GB from $1,656.15 a month, bare metal with root access and unlimited traffic.