Skip to main content

Documentation Index

Fetch the complete documentation index at: https://runcrate.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

Run ComfyUI — the node-based workflow editor for Stable Diffusion and FLUX — on a dedicated cloud GPU. Generate images faster than your local machine, run batch jobs 24/7, and access high-VRAM GPUs for large models like FLUX.1 Pro and SDXL.

​
What you’ll build

A remote ComfyUI instance accessible from your browser, running on a dedicated GPU with all models and custom nodes pre-installed. You can connect from anywhere and generate images without taxing your local hardware.

​
GPU sizing

ModelMin VRAMRecommended GPU
SDXL12 GBRTX 4090
FLUX.1 Schnell16 GBRTX 4090
FLUX.1 Dev / Pro24 GBRTX 4090, L40S
FLUX.2 Pro (4MP)32 GB+L40S, A100
Multiple models loaded48 GB+L40S, A100

​
Setup (CLI)

​
1. Deploy a GPU instance

# RTX 4090 for SDXL/FLUX.1, L40S for FLUX.2 or multiple models
runcrate instances create --name comfyui --gpu RTX4090

​
2. Install ComfyUI

runcrate ssh comfyui -- "
  cd /root &&
  git clone https://github.com/comfyanonymous/ComfyUI.git &&
  cd ComfyUI &&
  pip install -r requirements.txt
"

​
3. Download models

# FLUX.1 Schnell (fast, good quality)
runcrate ssh comfyui -- "
  cd /root/ComfyUI/models/checkpoints &&
  wget -q 'https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/flux1-schnell.safetensors'
"

# SDXL base
runcrate ssh comfyui -- "
  cd /root/ComfyUI/models/checkpoints &&
  wget -q 'https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors'
"

# CLIP and VAE for FLUX
runcrate ssh comfyui -- "
  cd /root/ComfyUI/models/clip &&
  wget -q 'https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp8_e4m3fn.safetensors' &&
  wget -q 'https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors'
"

runcrate ssh comfyui -- "
  cd /root/ComfyUI/models/vae &&
  wget -q 'https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors'
"

runcrate ssh comfyui -- "
  cd /root/ComfyUI/custom_nodes &&
  git clone https://github.com/ltdrdata/ComfyUI-Manager.git &&
  git clone https://github.com/Fannovel16/comfyui_controlnet_aux.git &&
  git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus.git &&
  cd comfyui_controlnet_aux && pip install -r requirements.txt
"

​
5. Start ComfyUI

runcrate ssh comfyui -- "
  cd /root/ComfyUI &&
  nohup python main.py --listen 0.0.0.0 --port 8188 > /root/comfyui.log 2>&1 &
"

​
6. Connect

# Get the IP
runcrate instances info comfyui
Open http://<INSTANCE_IP>:8188 in your browser. ComfyUI is now running on your cloud GPU.

​
Persistent setup with storage

Downloading models every time you deploy a new instance is slow. Use a storage volume to persist your models: 
# Create a storage volume for models
runcrate volumes create --name comfyui-models --size 200 --region us-east-1

# Deploy instance with storage attached
runcrate instances create --name comfyui --gpu RTX4090 --storage comfyui-models

# Symlink model directories to the persistent volume
runcrate ssh comfyui -- "
  mkdir -p /workspace/checkpoints /workspace/clip /workspace/vae /workspace/controlnet &&
  ln -sf /workspace/checkpoints /root/ComfyUI/models/checkpoints &&
  ln -sf /workspace/clip /root/ComfyUI/models/clip &&
  ln -sf /workspace/vae /root/ComfyUI/models/vae &&
  ln -sf /workspace/controlnet /root/ComfyUI/models/controlnet
"
# Download models to /workspace/checkpoints/ (they persist across instances)
Next time you deploy with the same volume, the models are already there — just symlink and start ComfyUI.

​
Using MCP

“Deploy an RTX 4090 instance called comfyui. Clone the ComfyUI repo, install requirements, download the FLUX.1 Schnell checkpoint, and start the server on port 8188. Give me the URL when it’s ready.”
The agent runs through the full setup and returns the URL to open in your browser.

​
Tips

  • Save workflows as JSON — export from ComfyUI and store in your repo for reproducible generation
  • Use FP8 quantized models for FLUX.2 — 40% less VRAM, similar quality
  • Kill the instance when done — runcrate instances delete comfyui stops billing immediately
  • Use the API mode — ComfyUI has a REST API at /prompt for programmatic batch generation