Workflows

cmfy.cloud executes ComfyUI workflows. This page explains the workflow format and how to prepare your workflows for the API.

What Is a Workflow?

A ComfyUI workflow is a graph of nodes that process data. Each node performs one operation:

In this example:

1. Load Checkpoint - Loads the AI model from disk
2. CLIP Encode - Converts your text prompt to embeddings
3. Empty Latent - Creates a blank latent image
4. KSampler - The main generation step
5. VAE Decode - Converts latent to pixel image
6. Save Image - Outputs the final image

Workflow Format

The API accepts workflows in ComfyUI's native format - a JSON object where each key is a node ID:

{
  "3": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "https://huggingface.co/stabilityai/sdxl/model.safetensors"
    }
  },
  "6": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "a beautiful sunset over mountains",
      "clip": ["3", 1]
    }
  },
  "5": {
    "class_type": "EmptyLatentImage",
    "inputs": {
      "width": 1024,
      "height": 1024,
      "batch_size": 1
    }
  }
}

Node Structure

Every node has:

Field	Required	Description
class_type	Yes	The ComfyUI node type (e.g., KSampler, VAEDecode)
inputs	Yes	Parameters for this node
_meta	No	Optional metadata (ignored by the system)

Connecting Nodes

Nodes connect by referencing other nodes' outputs. The format is [node_id, output_index]:

{
  "6": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "a cat",
      "clip": ["3", 1]  // Output #1 from node "3"
    }
  }
}

The ["3", 1] means "take output slot 1 from node 3". Output slots are zero-indexed, so slot 1 is the second output from that node (slot 0 is the first).

Converting from ComfyUI

If you have a workflow in the ComfyUI desktop app:

Step 1: Enable Dev Mode

In ComfyUI, go to Settings > Enable Dev mode options.

Step 2: Save API Format

Click Save (API Format) instead of regular Save. This exports the workflow in the format the API expects.

Step 3: Replace Model Paths

Change local file paths to URLs:

// Before (local file)
"ckpt_name": "v1-5-pruned-emaonly.ckpt"

// After (URL)
"ckpt_name": "https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.ckpt"

Model URLs

cmfy.cloud downloads models on-demand from URLs you provide. Supported sources:

Source	URL Pattern
Hugging Face	https://huggingface.co/...
Civitai	https://civitai.com/api/download/...
Amazon S3	https://*.s3.amazonaws.com/...
Google Cloud	https://storage.googleapis.com/...
Azure Blob	https://*.blob.core.windows.net/...
Cloudflare R2	https://*.r2.cloudflarestorage.com/...

Use Consistent URLs

Always use the same URL for the same model. The routing system uses URL matching to find nodes with cached models. Different URLs for the same model will be treated as different models.

Common Node Types

Here are frequently used nodes and their typical inputs:

Checkpoint Loaders

Load the main AI model:

{
  "class_type": "CheckpointLoaderSimple",
  "inputs": {
    "ckpt_name": "https://..."
  }
}

Text Encoders

Convert text to model-understandable embeddings:

{
  "class_type": "CLIPTextEncode",
  "inputs": {
    "text": "your prompt here",
    "clip": ["checkpoint_node", 1]
  }
}

Samplers

The core generation step:

{
  "class_type": "KSampler",
  "inputs": {
    "seed": 42,
    "steps": 20,
    "cfg": 7.5,
    "sampler_name": "euler",
    "scheduler": "normal",
    "denoise": 1.0,
    "model": ["checkpoint_node", 0],
    "positive": ["positive_prompt_node", 0],
    "negative": ["negative_prompt_node", 0],
    "latent_image": ["latent_node", 0]
  }
}

LoRA Loaders

Add style or character LoRAs:

{
  "class_type": "LoraLoader",
  "inputs": {
    "lora_name": "https://...",
    "strength_model": 0.8,
    "strength_clip": 0.8,
    "model": ["checkpoint_node", 0],
    "clip": ["checkpoint_node", 1]
  }
}

Workflow Limits

To ensure fair resource usage:

Limit	Value
Maximum nodes	500
Maximum payload size	10 MB
Required fields	class_type, inputs

Best Practices

1. Use Minimal Workflows

Include only the nodes you need. Smaller workflows are faster to validate and execute.

2. Reuse Model URLs

When possible, use popular models that other users also use. These are more likely to be cached on GPU nodes.

3. Test Locally First

Run your workflow in ComfyUI locally before submitting to the API. This catches errors faster.

4. Use Deterministic Seeds

Set explicit seed values for reproducible results. Random seeds make debugging harder.

Example: Complete Workflow

Here's a minimal text-to-image workflow:

{
  "1": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors"
    }
  },
  "2": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "a serene mountain landscape at sunset, digital art",
      "clip": ["1", 1]
    }
  },
  "3": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "text": "blurry, low quality",
      "clip": ["1", 1]
    }
  },
  "4": {
    "class_type": "EmptyLatentImage",
    "inputs": {
      "width": 1024,
      "height": 1024,
      "batch_size": 1
    }
  },
  "5": {
    "class_type": "KSampler",
    "inputs": {
      "seed": 12345,
      "steps": 25,
      "cfg": 7.0,
      "sampler_name": "euler_ancestral",
      "scheduler": "normal",
      "denoise": 1.0,
      "model": ["1", 0],
      "positive": ["2", 0],
      "negative": ["3", 0],
      "latent_image": ["4", 0]
    }
  },
  "6": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["5", 0],
      "vae": ["1", 2]
    }
  },
  "7": {
    "class_type": "SaveImage",
    "inputs": {
      "filename_prefix": "output",
      "images": ["6", 0]
    }
  }
}

What's Next?

• Cache-Aware Routing - Learn how model caching affects performance
• API Reference - Full API documentation