Yes. io.net offers one-click ComfyUI deployment with pre-configured GPU instances optimized for image generation workflows. Deploy ComfyUI on RTX 4090 for $0.18/hour or A100 for $1.10/hour, with all dependencies, models, and custom nodes pre-installed. Access your ComfyUI workspace via web browser with persistent storage for workflows and generated images.
ComfyUI on io.net supports all popular models (SDXL, SD 1.5, Flux, Pony Diffusion), custom nodes (ControlNet, IPAdapter, AnimateDiff), and high-resolution generation up to 4K. The platform handles automatic model downloading, GPU optimization, and workflow persistence—deploy in under 60 seconds.
Quick Start: Deploy ComfyUI
# Install io.net CLI
pip install ionet-cli
io login
# One-click ComfyUI deployment
io deploy comfyui \
--gpu RTX4090 \
--storage 200GB \
--name my-comfyui
# Returns URL in ~60 seconds:
# https://xxx-comfyui.ionet.cloud
# Access via browser, start generating immediately
Cost: $0.18/hour (RTX 4090) = ~$1.44 for 8-hour session
Pre-Configured Features
Included Models:
- Stable Diffusion XL (Base + Refiner)
- SD 1.5 (multiple checkpoints)
- Flux.1 [Dev]
- Pony Diffusion XL
- ControlNet models (Canny, Depth, OpenPose, etc.)
- VAE models (sdxl-vae-fp16-fix, vae-ft-mse-840000)
Pre-Installed Custom Nodes:
- ControlNet Preprocessors
- IPAdapter Plus
- AnimateDiff Evolved
- ComfyUI Manager (install additional nodes)
- Image Saver
- Efficiency Nodes
- UltimateSDUpscale
Storage:
- /workspace/ComfyUI/models: Pre-loaded models (~50GB)
- /workspace/ComfyUI/output: Generated images (persistent)
- /workspace/ComfyUI/input: Upload source images
- /workspace/ComfyUI/custom_nodes: Additional node installations
GPU Selection for ComfyUI
| GPU Type | Cost/Hour | Best For | Generation Speed (SDXL 1024x1024) |
|---|---|---|---|
| RTX 4090 | $0.18 | Cost-effective, excellent performance | 5-7 sec/image (25 steps) |
| RTX 3090 | $0.12 | Budget option, good for SD 1.5 | 8-12 sec/image |
| A100 | $1.10 | Batch generation, high-res (4K+) | 3-5 sec/image |
| L40S | $0.65 | Professional workflows, 48GB VRAM | 4-6 sec/image |
| H100 | $1.49-2.20 | Maximum speed, 8K generation | 2-3 sec/image |
Recommendation: RTX 4090 offers the best price/performance for most ComfyUI workflows.
Advanced Configuration
Custom Model Loading:
# Deploy with custom models from HuggingFace
io deploy comfyui \
--gpu RTX4090 \
--storage 500GB \
--env HUGGINGFACE_TOKEN=$HF_TOKEN \
--env MODELS="stabilityai/stable-diffusion-xl-base-1.0,runwayml/stable-diffusion-v1-5" \
--name custom-comfyui
# Or mount external storage with your model library
io deploy comfyui \
--gpu RTX4090 \
--mount s3://my-models:/workspace/ComfyUI/models \
--name comfyui-prod
Multi-GPU for Batch Generation:
# Deploy ComfyUI with 4x RTX 4090 for parallel batch processing
io deploy comfyui \
--gpu RTX4090 --count 4 \
--batch-mode \
--name comfyui-batch
# Automatically distributes queue across GPUs
# 4x speedup for batch jobs (100 images in 2 minutes vs. 8 minutes)
Custom Node Installation:
Access terminal via io.net dashboard or SSH:
# SSH into ComfyUI instance
io ssh my-comfyui
# Install ComfyUI Manager (if not pre-installed)
cd /workspace/ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git
# Install specific custom node
cd /workspace/ComfyUI/custom_nodes
git clone https://github.com/Fannovel16/comfyui_controlnet_aux.git
pip install -r comfyui_controlnet_aux/requirements.txt
# Restart ComfyUI
supervisorctl restart comfyui
Workflow Examples
Text-to-Image (SDXL):
{
"nodes": {
"1": {"type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": "sd_xl_base_1.0.safetensors"}},
"2": {"type": "CLIPTextEncode", "inputs": {"text": "epic fantasy landscape, detailed, 8k"}},
"3": {"type": "KSampler", "inputs": {"steps": 25, "cfg": 7.5}},
"4": {"type": "VAEDecode"},
"5": {"type": "SaveImage"}
}
}
ControlNet (Pose to Image):
{
"nodes": {
"1": {"type": "LoadImage", "inputs": {"image": "pose.png"}},
"2": {"type": "ControlNetLoader", "inputs": {"control_net_name": "control_v11p_sd15_openpose.pth"}},
"3": {"type": "CLIPTextEncode", "inputs": {"text": "professional photo, detailed"}},
"4": {"type": "ControlNetApply"},
"5": {"type": "KSampler", "inputs": {"steps": 30, "cfg": 8.0}},
"6": {"type": "SaveImage"}
}
}
High-Res Upscaling:
{
"nodes": {
"1": {"type": "LoadImage", "inputs": {"image": "generated_512x512.png"}},
"2": {"type": "UltimateSDUpscale", "inputs": {"upscale_by": 4, "tile_size": 512}},
"3": {"type": "SaveImage"}
}
}
Performance Benchmarks
SDXL Base (1024x1024, 25 steps):
| GPU | Time per Image | Batch 10 Images | Cost per 100 Images |
|---|---|---|---|
| RTX 4090 | 6 sec | 60 sec | $0.30 |
| A100 | 4 sec | 40 sec | $1.22 |
| RTX 3090 | 10 sec | 100 sec | $0.33 |
| H100 | 2.5 sec | 25 sec | $1.03 |
SDXL Refiner (1024x1024, 15 steps):
| GPU | Time per Image | Cost per 100 Images |
|---|---|---|
| RTX 4090 | 4 sec | $0.20 |
| A100 | 2.5 sec | $0.76 |
SD 1.5 (512x512, 20 steps):
| GPU | Time per Image | Batch 10 Images | Cost per 100 Images |
|---|---|---|---|
| RTX 4090 | 2 sec | 20 sec | $0.10 |
| RTX 3090 | 3 sec | 30 sec | $0.10 |
Use Cases and Workflows
1. Content Creation at Scale:
# Generate 1,000 product images
# Cost on AWS (g5.xlarge @ $1.21/hr): ~$20
# Cost on io.net (RTX 4090 @ $0.18/hr): ~$3
# Savings: 85%
2. Art Generation and Experimentation:
- Try multiple models and styles quickly
- Iterate on prompts without high costs
- Share workspace URL with team for collaboration
3. Commercial Image Generation:
- Generate marketing materials, social media content
- Batch process client requests
- High-res outputs for print (4K-8K with upscaling)
4. Animation with AnimateDiff:
# Deploy ComfyUI with AnimateDiff pre-configured
io deploy comfyui \
--gpu A100 \
--preset animatediff \
--storage 300GB
# Generate 16-frame animations (512x512)
# Time: ~45 seconds per animation on A100
API Integration
ComfyUI on io.net includes API access for programmatic generation:
import requests
import json
# ComfyUI API endpoint
url = "https://xxx-comfyui.ionet.cloud/prompt"
# Workflow JSON
workflow = {
"prompt": {
"3": {
"class_type": "KSampler",
"inputs": {
"seed": 42,
"steps": 25,
"cfg": 7.5,
"sampler_name": "euler_ancestral",
"scheduler": "normal",
"denoise": 1.0
}
},
"4": {"class_type": "CLIPTextEncode", "inputs": {"text": "beautiful landscape"}},
# ... rest of workflow
}
}
# Submit generation
response = requests.post(url, json=workflow)
prompt_id = response.json()['prompt_id']
# Poll for completion
while True:
status = requests.get(f"{url}/history/{prompt_id}").json()
if prompt_id in status and status[prompt_id].get('status', {}).get('completed'):
break
time.sleep(1)
# Download image
image_url = f"{url}/view?filename={status[prompt_id]['outputs']['9']['images'][0]['filename']}"
image = requests.get(image_url).content
Cost Comparison: io.net vs. Other Platforms
Scenario: Generate 500 SDXL images (1024x1024, 25 steps)
| Platform | GPU | Time | Cost | Notes |
|---|---|---|---|---|
| io.net | RTX 4090 | 50 min | $0.15 | Best value |
| RunPod | RTX 4090 | 50 min | $0.28 | 87% more expensive |
| AWS (g5.xlarge) | A10G | 90 min | $1.82 | 1,113% more expensive |
| Replicate | A100 | 40 min | $3.50 | 2,233% more expensive (inference API) |
| Local (owned RTX 4090) | RTX 4090 | 50 min | $0 | Requires $1,600 GPU purchase |
Break-even: Need to generate 533 hours of images locally to match owning RTX 4090 vs. renting on io.net ($1,600 / $0.18/hr = 8,889 hrs of use)
Persistent Workflows and Model Storage
# Save workspace state
# All workflows, generated images, and custom nodes persist in storage
# Stop instance when not in use
io stop my-comfyui
# Restart later with all data intact
io start my-comfyui
# Same URL, all workflows and models preserved
# Download generated images
io download my-comfyui:/workspace/ComfyUI/output ./local_images/
Troubleshooting
Out of VRAM:
- Reduce batch size in KSampler
- Lower resolution (2048x2048 → 1024x1024)
- Enable VAE tiling in UltimateSDUpscale
- Use FP16 models instead of FP32
- Upgrade to A100 (80GB VRAM) for 4K+ generation
Slow Generation:
- Check GPU utilization: nvidia-smi via SSH
- Ensure xformers is enabled (pre-configured on io.net)
- Consider upgrading to A100 or H100
- Use Euler Ancestral sampler (faster than DPM++)
Model Download Failures:
- Check storage space: df -h /workspace
- Increase storage: io resize-storage my-comfyui 500GB
- Use model cache: Mount S3 bucket with pre-downloaded models
Start generating images with ComfyUI on io.net — one-click deployment, $0.18/hour, RTX 4090 performance.
