Creating Your Own ComfyUI Workflows
Step-by-step guide to creating your own ComfyUI workflows – from simple to complex.
Table of Contents
01Your First Workflow
A minimal text-to-image workflow in ComfyUI consists of just 5 nodes: Load Checkpoint → CLIP Text Encode (Prompt) → CLIP Text Encode (Negative) → KSampler → VAE Decode → Save Image. This basic framework is the foundation for all more complex workflows.
02Step 1: Load Model
Add a 'Load Checkpoint' node (right-click → Add Node → Loaders). Select your model in the dropdown. The node has three outputs: MODEL (the neural network), CLIP (the text encoder), and VAE (the image decoder). Connect each output to the corresponding input of the following nodes.
03Step 2: Create Prompts
Add two 'CLIP Text Encode' nodes. Connect the CLIP output of the checkpoint loader to both. Use the first for your positive prompt (what you want to generate), and the second for the negative prompt (what should be avoided). Enter your prompts in the text fields.
04Step 3: Configure Sampling
The KSampler node is the heart of the workflow. Connect the MODEL output to the 'model' input, and the two CLIP encode outputs to 'positive' and 'negative'. Important settings:
- Steps: 20–30 for good quality (more steps = better but slower)
- CFG: 7–8 for natural images, 10+ for stronger prompt fidelity
- Sampler: 'euler_ancestral' for creative results, 'dpmpp_2m' for consistent quality
- Scheduler: 'karras' is a good default
- Seed: Fixed value for reproducible results, -1 for random
- Denoise: 1.0 for text-to-image, 0.3–0.7 for image-to-image
05Step 4: Output Image
Connect the KSampler output to a 'VAE Decode' node (which converts the latent image into pixels), and its output to a 'Save Image' or 'Preview Image' node. Click 'Queue Prompt' – and your first image will be generated!
06Extending the Workflow
From here, you can extend the workflow in any way:
- Add LoRA: 'Load LoRA' node between Checkpoint and KSampler
- ControlNet: 'Apply ControlNet' node for image guidance
- Upscaling: 'Upscale Image' node after VAE Decode for higher resolution
- img2img: 'Load Image' node as additional input for the KSampler
- Inpainting: Re-generate masked areas
- Batch Processing: Latent Batch node for multiple images at once
07Best Practices
Keep workflows organized with groups (right-click → Add Group). Name nodes descriptively with right-click → Title. Save working workflows regularly as JSON. Use 'Reroute' nodes for clean connections. And most importantly: experiment! ComfyUI rewards curiosity.