comfy-runtime

Run any ComfyUI node as a Python function. No server, no UI, no workflow graph.

Install

pip install comfy-runtime

Usage

1. Configure & run built-in nodes

import comfy_runtime

comfy_runtime.configure(models_dir="/path/to/models", output_dir="./output")

# Text-to-image with SD 1.5
model, clip, vae = comfy_runtime.execute_node("CheckpointLoaderSimple", ckpt_name="v1-5-pruned-emaonly.safetensors")
positive = comfy_runtime.execute_node("CLIPTextEncode", clip=clip, text="a castle on a hill, fantasy art")[0]
negative = comfy_runtime.execute_node("CLIPTextEncode", clip=clip, text="blurry, low quality")[0]
latent = comfy_runtime.execute_node("EmptyLatentImage", width=512, height=512, batch_size=1)[0]
sampled = comfy_runtime.execute_node("KSampler", model=model, positive=positive, negative=negative, latent_image=latent, seed=42, steps=20, cfg=8.0, sampler_name="euler", scheduler="normal", denoise=1.0)[0]
images = comfy_runtime.execute_node("VAEDecode", samples=sampled, vae=vae)[0]
comfy_runtime.execute_node("SaveImage", images=images, filename_prefix="output")

2. Load any custom node

# Single file
comfy_runtime.load_nodes_from_path("/path/to/my_node.py")

# Directory (with or without __init__.py)
comfy_runtime.load_nodes_from_path("/path/to/ComfyUI-AnimateDiff-Evolved")

# Then use it
result = comfy_runtime.execute_node("ADE_AnimateDiffLoRALoader", ...)

Both V1 (NODE_CLASS_MAPPINGS) and V3 (comfy_entrypoint) node formats are supported.

3. Discover nodes

comfy_runtime.list_nodes()          # All registered node names
comfy_runtime.get_node_info("KSampler")  # Input/output types, category

API

Custom node compatibility

Tested with popular third-party nodes:

Custom nodes that import from comfy.*, comfy_api.*, comfy_extras.*, folder_paths, or nodes all resolve correctly.

Examples

See workflows/ for complete runnable examples:

• sd15_text_to_image — basic text-to-image
• flux2_klein_text_to_image — Flux.2 with custom sampler
• img2img — image-to-image style transfer
• inpainting — masked region replacement
• hires_fix — 2-pass high-resolution generation
• area_composition — spatial prompt composition
• esrgan_upscale — 4x super-resolution

License

MIT