workflows

Files (1)

• SKILL.md

SKILL.md (first lines)

# ComfyUI Workflow Building Guide

## API Format

ComfyUI workflows use a JSON format where each node is a key-value pair:

```json
{
  "1": {
    "class_type": "KSampler",
    "inputs": {
      "seed": 42,
      "steps": 20,
      "cfg": 7.0,
      "sampler_name": "euler",
      "scheduler": "normal",
      "denoise": 1.0,
      "model": ["2", 0],
      "positive": ["3", 0],
      "negative": ["4", 0],
      "latent_image": ["5", 0]
    }
  }
}
```

- **Keys** are string node IDs (e.g., `"1"`, `"2"`)
- **`class_type`** is the ComfyUI node type name
- **`inputs`** contains parameters and connections
- **Connections** use the format `["source_node_id", output_index]`

## Common Node Chains

### txt2img (text to image)
`CheckpointLoaderSimple` -> `CLIPTextEncode` (positive) -> `KSampler` -> `VAEDecode` -> `SaveImage`

### img2img (image to image)
`LoadImage` -> `VAEEncode` -> `KSampler` (with denoise < 1.0) -> `VAEDecode` -> `SaveImage`

### ControlNet
`LoadImage` -> `ControlNetLoader` + `ControlNetApply` -> feed into `KSampler` positive conditioning

### LoRA Stacking
`CheckpointLoaderSimple` -> `LoraLoader` (chain multiple) -> `KSampler`

## Key Nodes Reference

| Node | Purpose | Key Inputs |
|------|---------|------------|
| `CheckpointLoaderSimple` | Load a model checkpoint | `ckpt_name` |
| `CLIPTextEncode` | Encode text prompt | `text`, `clip` |
| `KSampler` | Run diffusion sampling | `model`, `positive`, `negative`, `latent_image`, `steps`, `cfg`, `seed` |
| `VAEDecode` | Decode latent to image | `samples`, `vae` |
| `VAEEncode` | Encode image to latent | `pixels`, `vae` |
| `SaveImage` | Save output image | `images`, `filename_prefix` |
| `LoadImage` | Load input image | `image` (filename) |
| `LoraLoader` | Apply LoRA weights | `model`, `clip`, `lora_name`, `strength_model`, `strength_clip` |
| `ControlNetLoader` | Load ControlNet model | `control_net_name` |
| `ControlNetApply` | Apply ControlNet | `conditioning`, `control_net`, `image`, `strength` |

## Using the Workflow Tools

**Quick decision guide** — pick the tool by what you need to do:

| Goal | Tool |
|------|------|
| Build a new workflow from scratch | `comfyui_create_workflow` |
| Modify an existing workflow | `comfyui_modify_workflow` |
| Understand a workflow (human-readable text or Mermaid) | `comfyui_summarize_workflow` |
| Understand a workflow (machine-readable dict) | `comfyui_analyze_workflow` |
| Check a workflow for errors before running | `comfyui_validate_workflow` |
| Execute a workflow | `comfyui_run_workflow` |

1. **`comfyui_create_workflow`** — Start from a built-in template. Accepted template names are listed in the tool's docstring: `txt2img`, `img2img`, `upscale`, `inpaint`, `txt2vid_animatediff`, `txt2vid_wan`, `controlnet_canny`, `controlnet_depth`, `controlnet_openpose`, `ip_adapter`, `lora_stack`, `face_restore`, `flux_txt2img`, `sdxl_txt2img`. The `params` argument defaults to `""`, which applies the predefined settings for the selected template (e.g., default dimensions, steps, and cfg values baked into that template); pass a JSON string when you want overrides. (Note: `comfyui_list_workflows` returns server-side `/workflow_templates` from front-end packages — a different set, not these MCP-built-in templates.)
2. **`comfyui_modify_workflow`** — Add/remove nodes, change connections, update parameters on an existing workflow. Operations: `add_node`, `remove_node`, `set_input`, `connect`, `disconnect`. The docstring spells out each op's exact field shape. The call is atomic — if any operation fails, the call raises and the input workflow is unmodified.
3. **`comfyui_validate_workflow`** — Returns `{valid, errors, warnings, node_count, pipeline}`. `errors` (list[str]) are blocking; `warnings` (list[str]) are non-blocking concerns like missing model files or suspicious inputs.
4. **`comfyui_summarize_workflow`** — Human-readable overview. Pass `output_format="text"` (default) or `"mermaid"` for diagram markup.
5. **`comfyui_analyze_workflow`** — Returns the analysis as a structured dict (`node_count`, `class_types`, `flow`, `models`, `parameters`, `pipeline`, `prompt_nodes`, `negative_nodes`). Use this when you want to read fields like `pipeline` programmatically; use `comfyui_summarize_workflow` for the human-readable rendering.
6. **`comfyui_run_workflow`** — Submit a workflow for execution. Use `wait=True` to block until done. The response is always a dict envelope: `{status, prompt_id, outputs, elapsed_seconds, warnings?}` where `status` is one of `submitted` / `completed` / `interrupted` / `error` / `timeout`. With `wait=False` you get just `{status: "submitted", prompt_id, warnings?}` — poll `comfyui_get_progress` or `comfyui_get_job` to follow up.

## Tips

- Always validate a workflow before running it.
- Use `comfyui_list_nodes` to check what node types are available on the connected ComfyUI instance.
- Use `comfyui_get_node_info` to get detailed input/output specs for a specific node type.
- Node IDs must be unique strings. When adding nodes, pick IDs that don't conflict with existing ones.
- The `seed` parameter controls reproducibility. Use a fixed seed for consistent results, or -1 for random.

From hybridindie/comfyui_mcp, MIT.

More from hybridindie/comfyui_mcp