Stable Diffusion Server — ComfyUI & Flux on a Dedicated GPU
Rent a dedicated GPU server for Stable Diffusion, ComfyUI and Flux — a whole RTX 4090 or RTX 5090 in Frankfurt, with full VRAM for high-resolution image generation. A fixed monthly price, no queue, GDPR-compliant.
Request a GPU server for Stable DiffusionFree web UIs for Stable Diffusion are handy for first experiments — but for serious work they hit limits fast: queues, caps, shared GPUs with preemption, and images processed on someone else's servers outside the EU. A dedicated GPU server gives you full control instead: your own ComfyUI or Automatic1111 instance, the whole GPU to yourself, and a fixed monthly price with no per-second billing.
Why a dedicated server for Stable Diffusion?
Image generation is VRAM- and compute-heavy, especially with SDXL, Flux and high-resolution upscales. On shared or hourly cloud GPUs you compete for resources, wait in queues and risk aborted renders from preemption . At Bthorio your stack runs on dedicated bare metal — the whole RTX 4090 or RTX 5090 is yours, around the clock, at a fixed price.
How much VRAM do SDXL and Flux need?
Whether a workflow runs smoothly depends mostly on available VRAM . SD 1.5 gets by on a few gigabytes, SDXL wants around 10–12 GB, and Flux plus video models can ask for considerably more. The table below shows what fits on a 24 GB RTX 4090 or a 32 GB RTX 5090.
| Model / workflow | VRAM needed (approx.) | RTX 4090 (24 GB) | RTX 5090 (32 GB) |
|---|---|---|---|
| SD 1.5 (512 px) | ~4 GB | Effortless | Effortless |
| SDXL (1024 px) | ~10–12 GB | Yes | Yes |
| SDXL + ControlNet / LoRA | ~14–18 GB | Yes | Yes, with headroom |
| Flux.1 [dev] (fp8) | ~16–20 GB | Yes | Yes |
| Flux.1 [dev] (fp16/bf16) | ~24 GB+ | Tight | Yes |
| Video (AnimateDiff / SVD) | ~20–30 GB+ | Limited | Yes |
These are guide values including overhead for the VAE, ControlNet and upscalers; very large batches, high resolutions or several concurrent pipelines push the requirement up. For most SDXL and Flux workflows the RTX 4090 is comfortable; if you constantly work at the VRAM ceiling or generate video, reach for the RTX 5090.
ComfyUI, Automatic1111 or Forge — your choice
- ComfyUI — node-based pipelines for full control over Flux and SDXL workflows
- Automatic1111 / Forge — the classic web UI for fast prompting and extensions
- Your own ControlNet, LoRA and IP-Adapter models with no upload limits
- API access for automated image pipelines and batch generation
- Upscaling, inpainting and face restore in the same instance, no separate services
RTX 4090 or RTX 5090?
For SDXL, most Flux workflows and ComfyUI pipelines the RTX 4090 with 24 GB from €399/month is more than enough. Once you run Flux at full precision, video generation or very large batches, the RTX 5090 with 32 GB gives you the headroom you need. You'll find both on our rent a GPU server overview.
A fixed monthly price instead of per-second billing
With hourly or per-second cloud GPUs you pay for every second of compute — and long render sessions, batch jobs or experiments drive costs up unpredictably. A dedicated server costs the same every month, no matter how many images you generate. For creative studios, agencies and shops with ongoing image needs that's planable rather than risky.
Up and running in minutes
You have root access and set up ComfyUI or Automatic1111 in minutes — you choose the operating system, drivers and CUDA version freely. For a full step-by-step walkthrough see our guide to setting up Stable Diffusion — from driver and CUDA through installation to your first image.