Last Updated: December 24, 2025
Essential Node Types
| Node | Purpose | Category |
|---|---|---|
| CheckpointLoaderSimple | Load base model (SD 1.5, SDXL, etc.) | Loaders |
| CLIPTextEncode | Process text prompts (positive/negative) | Conditioning |
| KSampler | Generate images from noise | Sampling |
| VAEDecode | Convert latent to pixel space | Latent |
| SaveImage | Save output images | Image |
| LoadImage | Load reference images | Image |
| LoraLoader | Add LoRA modifications | Loaders |
| ControlNetLoader | Load ControlNet models | Loaders |
Basic Text-to-Image Workflow
1. CheckpointLoaderSimple
├─→ MODEL → KSampler
├─→ CLIP → CLIPTextEncode (Positive) → KSampler
├─→ CLIP → CLIPTextEncode (Negative) → KSampler
└─→ VAE → VAEDecode
2. KSampler
└─→ LATENT → VAEDecode
3. VAEDecode
└─→ IMAGE → SaveImage
# Connections Flow:
Checkpoint → CLIP → Text Encode → Sampler → VAE Decode → SaveKSampler Parameters
| Parameter | Range | Description |
|---|---|---|
| seed | 0 - 2^32 | Random seed for reproducibility |
| steps | 1 - 150 | Sampling steps (20-30 typical) |
| cfg | 1.0 - 30.0 | Classifier-free guidance (7-11 typical) |
| sampler_name | Various | Euler, DPM++, etc. |
| scheduler | Various | Normal, Karras, Exponential |
| denoise | 0.0 - 1.0 | Denoising strength (1.0 = full) |
Popular Samplers
Euler a
Fast, good for quick iterations (ancestral)
DPM++ 2M Karras
High quality, balanced speed
DPM++ SDE Karras
Very high quality, slower
UniPC
Fast convergence, fewer steps needed
DDIM
Deterministic, good for img2img
Image-to-Image Workflow
1. LoadImage
└─→ IMAGE → VAEEncode
2. VAEEncode
└─→ LATENT → KSampler
3. KSampler (set denoise 0.4-0.8)
└─→ LATENT → VAEDecode
4. VAEDecode
└─→ IMAGE → SaveImage
# Key: Lower denoise = closer to original
# 0.3-0.5: Small changes
# 0.6-0.8: Moderate changes
# 0.9-1.0: Major changesLoRA Integration
1. CheckpointLoaderSimple
├─→ MODEL → LoraLoader
└─→ CLIP → LoraLoader
2. LoraLoader
├─→ MODEL → KSampler
├─→ CLIP → CLIPTextEncode
# strength_model: 0.5-1.0 (typical 0.7-0.9)
# strength_clip: 0.5-1.0 (typically same as model)
# Multiple LoRAs: Chain LoraLoaders
LoraLoader1 → LoraLoader2 → LoraLoader3 → KSamplerControlNet Workflow
1. LoadImage (reference)
└─→ IMAGE → ControlNetPreprocessor
2. ControlNetPreprocessor (Canny, Depth, etc.)
└─→ IMAGE → ControlNetApply
3. ControlNetLoader
└─→ CONTROL_NET → ControlNetApply
4. ControlNetApply
├─→ CONDITIONING (to KSampler positive)
# strength: 0.5-1.5 (control influence)
# Common Preprocessors:
- CannyEdgePreprocessor: Edge detection
- DepthEstimation: Depth maps
- OpenPosePreprocessor: Pose detection
- LineArtPreprocessor: Line art conversionUpscaling Workflow
# Method 1: Latent Upscale
1. KSampler → LATENT → LatentUpscale
2. LatentUpscale → LATENT → KSampler (high-res pass)
3. KSampler → VAEDecode → SaveImage
# upscale_method: nearest-exact, bilinear, area
# denoise: 0.4-0.6 for refinement
# Method 2: Ultimate SD Upscale
1. Generate base image
2. UpscaleModelLoader (RealESRGAN, etc.)
3. ImageUpscaleWithModel
4. VAEEncode → KSampler (denoise 0.3-0.5)
5. VAEDecode → SaveImageBatch Processing
# Batch Seeds
1. Add "Seed" node set to randomize/increment
2. Set batch_size in Empty Latent Image
3. SaveImage will output multiple files
# Batch Images
1. LoadImage → batch_size parameter
2. Process through workflow
3. Multiple outputs automatically saved
# Prompt Batch
1. Use "String" nodes with different prompts
2. "Primitive" node for switching
3. Queue multiple times with different valuesCustom Node Essentials
ComfyUI-Manager
Install and manage custom nodes easily
WAS Node Suite
Image processing and utilities
Efficiency Nodes
Streamlined workflow nodes
Impact Pack
Advanced sampling and detailing
ControlNet Preprocessors
All ControlNet preprocessors
Keyboard Shortcuts
| Shortcut | Action |
|---|---|
Ctrl + Enter |
Queue prompt |
Ctrl + Shift + Enter |
Queue prompt (front of queue) |
Space |
Add node menu |
Ctrl + C / V |
Copy/paste nodes |
Ctrl + M |
Mute/unmute node |
Ctrl + B |
Bypass node |
Delete |
Delete selected nodes |
Ctrl + G |
Group selected nodes |
Workflow Optimization
Use lower steps for testing (15-20)
Faster iterations during development
Cache models in memory
Enable in settings for faster switching
Use Efficiency Nodes
Combine multiple nodes into one
Bypass unused nodes
Ctrl+B to skip processing
Save workflows as .json
Right-click → Save Workflow
Troubleshooting Common Issues
| Issue | Solution |
|---|---|
| Out of memory | Lower resolution, reduce batch size, use --lowvram flag |
| Nodes not connecting | Check data type compatibility (MODEL, CLIP, LATENT, etc.) |
| Blurry output | Check VAE settings, try different VAE models |
| Slow generation | Use faster samplers, reduce steps, check GPU usage |
| Missing models | Place in correct folders: models/checkpoints, models/loras, etc. |
Model Folder Structure
ComfyUI/
├── models/
│ ├── checkpoints/ # .safetensors base models
│ ├── loras/ # LoRA files
│ ├── vae/ # VAE models
│ ├── controlnet/ # ControlNet models
│ ├── upscale_models/ # Upscalers (ESRGAN, etc.)
│ ├── clip/ # CLIP models
│ └── embeddings/ # Textual inversions
├── custom_nodes/ # Custom node extensions
└── output/ # Generated imagesAdvanced: API Usage
import json
import requests
# Load workflow JSON
with open('workflow.json') as f:
workflow = json.load(f)
# Modify prompt
workflow['6']['inputs']['text'] = "your new prompt here"
# Queue prompt
response = requests.post(
'http://127.0.0.1:8188/prompt',
json={'prompt': workflow}
)
# Get result
prompt_id = response.json()['prompt_id']
# Poll for completion and download images
💡 Pro Tip: Start with the default workflow and modify it incrementally. Save variations with descriptive names like "txt2img_basic.json", "controlnet_canny.json". Use the Manager extension to easily install custom nodes and keep everything updated.