6 min read
Stable Diffusion and FLUX on a rented GPU: what to choose and how to start
Which card to pick for SD 1.5, SDXL and FLUX, how to start Forge or ComfyUI on a rented server, and how to keep costs under control.
Generating images on a rented server means paying only for the minutes you work, not for an idle card. The first question is memory: it decides which models will run at all.
Which card you need
- SD 1.5: from 8 GB. Fast images and a huge LoRA library.
- SDXL: from 12 GB, comfortable at 16 GB.
- FLUX.1 in fp8: from 16 GB. FLUX.1 in bf16: from 24 GB.
In the catalog, filter cards by memory and compare the hourly price. For most SDXL and FLUX work a 16–24 GB card is enough.
Running in Forge
- Rent a server with the Forge template and the memory you need.
- Open the web interface from the server card.
- Put checkpoints in models/Stable-diffusion and LoRAs in models/Lora. Large files are easier to upload with SFTP or wget on the server.
cd models/Stable-diffusion
wget -O flux1-dev-fp8.safetensors "<direct file link>"Running in ComfyUI
ComfyUI suits FLUX and video models: a workflow is a graph you can save to a file. Put the model in models/checkpoints and load a ready workflow from its file.
Costs
Time one generation at your resolution and step count, then multiply by the per-minute GPU price. Stop the server while you tune prompts: a stopped server does not use GPU minutes. The disk is still billed, so keep only what you need on it.