English | 日本語 | 中文 | Español | Français | हिन्दी | Italiano | Português (BR)

Drive ComfyUI from Python. No node graph.

---

What this is

comfy-headless builds ComfyUI API-format graphs and runs them. You call a Python function; it emits the JSON node graph, POSTs it to ComfyUI, polls for completion, and hands you the output paths.

That framing matters, because it tells you what can go wrong. The library's entire job is emitting node names and input keys that the target ComfyUI actually has. When ComfyUI renames or drops a node, a graph referencing the old name is rejected at submit with an opaque error — and nothing warns you first.

v3.0 is the release that took that seriously. Every node type this library emits was audited against the live ComfyUI catalog. Nine no longer existed. They're gone, the graphs that used them are rebuilt on verified nodes, and the library can now tell you what a server is missing before you spend a run on it.

v3.1 extends that discipline to six workflow profiles — Image, Video, 3D, Inference, Metadata, Audio — on a shared addressing/typing layer that speaks ComfyUI's own validation rules: union type matching transcribed from the server's validator, dotted dynamic-combo fields (codec.encoding.crf), and conditional inputs that refuse inactive-branch values at construction time. Same route surface as before — meshes, music and captions come back through /history + /view like everything else.

Problem	What comfy-headless does
The node interface is a lot	Presets and a clean Python API
Prompt engineering is hard	Optional AI enhancement via local Ollama
Video generation is fiddly	26 presets across 9 model families
"I need a mesh from this image"	generate_3d() — Hunyuan3D-2, all core nodes
"I need music / stems"	generate_audio() (ACE-Step 1.5), separate_audio()
"What's in this image?"	run_inference() — caption, tag, detect, segment, OCR
"Which graph made this PNG?"	extract_prompt_graph() / rerun_from_png()
"Which settings do I use?"	Recommendations sized to your VRAM
Graphs fail with cryptic errors	Dependency check names the node and the pack

Quick start

pip install comfy-headless[standard]

from comfy_headless import ComfyClient

client = ComfyClient()                       # defaults to http://localhost:8188
result = client.generate_image("a beautiful sunset over mountains")
print(result["images"])

generate_image returns a dict with success, prompt_id, images, error, seed and preset. Every generation call in this library returns a dict — there is no result object to unwrap.

Install

pip install comfy-headless              # core only, ~2MB
pip install comfy-headless[standard]    # + AI enhancement + WebSocket (recommended)
pip install comfy-headless[full]        # everything

Extra	Adds
ai	Prompt analysis and enhancement via local Ollama
websocket	Real-time progress over WebSocket
ui	Gradio web interface
health	System health monitoring
validation	Pydantic config validation
observability	OpenTelemetry tracing
standard	ai + websocket
full	All of the above

Requires Python 3.10+ and a running ComfyUI instance.

Check what's active at runtime:

from comfy_headless import FEATURES, list_missing_features

print(FEATURES)                 # {'ai': True, 'websocket': True, 'health': False, ...}
print(list_missing_features())  # {'health': 'pip install comfy-headless[health]', ...}

Images

result = client.generate_image(
    "a cyberpunk street at night",
    negative_prompt="blurry, low quality",
    preset="hd",          # overrides width/height/steps/cfg when set
    seed=42,
)

Eight image presets: draft, fast, quality, hd, portrait, landscape, cinematic, square.

Batch a list of prompts:

result = client.generate_batch(
    ["a red fox", "a snowy owl", "a grey wolf"],
    preset="fast",
)

AI prompt enhancement

Requires the [ai] extra and a local Ollama. These are module-level functions, not client methods:

from comfy_headless import enhance_prompt, analyze_prompt

result = enhance_prompt("a cat", style="balanced")
print(result.enhanced)   # .original .enhanced .negative .additions .reasoning

analysis = analyze_prompt("a cyberpunk city at night")
print(analysis.intent, analysis.styles, analysis.suggested_preset)

Qwen-Image (new in 3.1)

Qwen-Image-2512 text-to-image ships as the qwen_txt2img template, with the recipe the model actually wants baked in: UNETLoader path, 16-channel EmptySD3LatentImage (the SDXL latent produces garbage on DiT models), steps 20, cfg 2.5, shift 3.1, native 1328×1328 bucket:

from comfy_headless import compile_workflow

compiled = compile_workflow("a castle above the clouds", template_id="qwen_txt2img")
prompt_id = client.queue_prompt(compiled.workflow)

Instruction editing with up to three reference images (Qwen-Image-Edit-2511):

result = client.edit_image(
    "make it night, keep the composition",
    images=["photo.png"],          # local paths, bytes, or uploaded refs — 1 to 3
)

References feed TextEncodeQwenImageEditPlus as discrete image1..image3 inputs and do not pass through VAEEncode — the graph shape the node actually expects.

ControlNet (new in 3.1)

One code path covers Qwen and SDXL union ControlNets:

from comfy_headless import build_controlnet_workflow

workflow = build_controlnet_workflow(
    "a stone fortress at dawn",
    control_image_ref=client.upload_image("depth.png")["ref"],
    control_type="depth",      # verbatim enum; "auto" makes the model infer
    base="qwen",               # or "sdxl"
)
client.queue_prompt(workflow)

Only the core Canny preprocessor is emitted (preprocess="canny"); other hint types expect a pre-made control image, because their preprocessors live in a custom pack this library does not silently require.

Video

from comfy_headless import list_video_presets, get_recommended_preset

print(list_video_presets())                  # 26 presets
print(get_recommended_preset(vram_gb=16))    # picks one that fits

result = client.generate_video(
    "a slow pan across a mountain range",
    preset="ltx_quality",
)
print(result["videos"])

Selection is by preset, not by model — generate_video has no model argument. Any of frames, fps, steps, cfg, width, height can be passed to override the preset.

Image input

Image-to-video, and anything else taking a source image, needs that image to exist inside ComfyUI first. Upload it, then pass the name the server gives back:

ref = client.upload_image("reference.png")
# {"name": "reference.png", "subfolder": "", "type": "input", "ref": "reference.png"}

result = client.generate_video(
    "a cat walking through a garden",
    preset="wan_14b",
    init_image=ref["name"],
)

Read name back from the response rather than reusing the filename you sent — ComfyUI renames on collision, so the two are not always equal. ref is the same value already joined with any subfolder, which is exactly what the graph needs.

Changed in 3.0: init_image is a server-side filename. Earlier versions accepted base64 data and smuggled it through a third-party node that does not exist on a stock ComfyUI install. See the CHANGELOG.

New in 3.1: Hunyuan 1.5 image-to-video is real i2v — presets hunyuan15_i2v and hunyuan15_i2v_fast build on the core HunyuanVideo15ImageToVideo node and require an init_image (v3.0 silently built a text-to-video graph instead). And output="core" swaps the VHS_VideoCombine terminator for core CreateVideo → SaveVideo, dropping the Video Helper Suite dependency entirely.

Model families

Family	Min VRAM	Quality	Speed	Extra nodes	Best for
Wan 2.1/2.2	6 GB	Great	Medium	—	Low VRAM, efficiency
AnimateDiff Lightning	6 GB	Fair	Fastest	AnimateDiff-Evolved	4-step drafts
AnimateDiff	8 GB	Good	Fast	AnimateDiff-Evolved, Frame Interp.	Quick previews
LTX-Video	12 GB	Excellent	Fast	—	The safe default
Mochi	12 GB	Excellent	Slow	—	Text adherence, long clips
SVD	12 GB	Good	Medium	—	Animating a still
Hunyuan 1.5	14 GB	Best	Slow	—	Highest quality
CogVideoX	16 GB	Good	Slow	CogVideoX Wrapper	Legacy
Hunyuan 1.0	24 GB	Great	Slow	Frame Interpolation	Superseded by 1.5

Six of the nine families run on stock ComfyUI core nodes — no wrapper pack for the model itself. Only AnimateDiff (both variants) and CogVideoX need one. Video output uses Video Helper Suite; frame interpolation uses Frame Interpolation.

VRAM figures are the floor for that family's default resolution, read from VIDEO_MODEL_INFO — not ceilings. More memory buys longer clips and higher resolutions from the same family.

Check before you run

Rather than discovering a missing node at submit time:

workflow = client.build_video_workflow("a cat walking")

report = client.check_workflow_dependencies(workflow)
print(report["missing_packs"])     # what this server is missing
print(report["required_packs"])    # what this graph needs

# or raise MissingNodePackError, naming the class and the pack that provides it
client.require_workflow_dependencies(workflow)

You can also type-check a graph's edges against the live server, using the server's own acceptance rule (so a MESH feeding a FILE_3D_* union input is never a false rejection):

report = client.check_workflow_types(workflow)
print(report["errors"])     # edges the server would reject
print(report["warnings"])   # accepted edges with partial type overlap

3D (new in 3.1)

Image-to-mesh via Hunyuan3D-2 — entirely ComfyUI core nodes, no wrapper packs, no new routes. The GLB registers in /history exactly like a PNG does and downloads through /view:

result = client.generate_3d("character.png", preset="detail")
# presets: standard / draft / detail
glb = client.get_file(**result["meshes"][0])
open("character.glb", "wb").write(glb)

generate_3d accepts a local path, raw bytes, an upload_image() dict, or a server-side ref, and uploads automatically when needed. It is pure image-conditioning — there is no text prompt in the graph. Tunables: steps (30), cfg (5.5), octree_resolution (256), threshold (0.6), seed.

Wrapper-pack 3D models (TRELLIS, TripoSG, ...) are deliberately not emitted — ComfyUI-3D-Pack's native dependencies are the least stable in the ecosystem.

Audio (new in 3.1)

Text-to-music via ACE-Step 1.5 — MIT-licensed code and weights, native core nodes, zero packs. The turbo checkpoint runs at 8 steps / cfg 1:

result = client.generate_audio(
    tags="lo-fi, jazz, mellow, rainy night",
    lyrics="",                       # empty = instrumental
    preset="music",                  # music / music_long / jingle / music_mp3 / draft
    seconds=30,
)
flac = client.get_file(**result["audios"][0])

The builder enforces the model's coupling invariant for you: the encoder's duration and the latent's seconds are one logical parameter, driven from a single field — the runtime does not cross-validate them, and a mismatch completes "successfully" with silently wrong output. Output goes through SaveAudioAdvanced (the only non-deprecated audio save node); flac output emits no quality field at all, mp3/opus emit the dotted format.quality sub-field.

Stem separation (requires the audio-separation-nodes-comfyui pack):

result = client.separate_audio("song.flac")            # bass, drums, other, vocals
result = client.separate_audio("song.flac", stems=["vocals"])

Inference (new in 3.1)

Non-generative model calls — ask questions about an image instead of making one. Runs on Florence-2 (pack comfyui-florence2; the detect task adds comfyui-segment-anything-2):

r = client.run_inference("photo.png", task="caption")
print(r["text"])                     # the caption, read straight from /history

r = client.run_inference("photo.png", task="tag")               # booru-style tags
r = client.run_inference("photo.png", task="ocr")
r = client.run_inference("photo.png", task="detect", text_input="the red car")
print(r["text"])                     # bounding-box coordinates as JSON

r = client.run_inference("photo.png", task="segment", text_input="the person")
mask_png = client.get_file(**r["images"][0])

The profile's load-bearing rule: a result reaches /history only through an output node. Every inference graph terminates in core SaveText (which reports the text inline — no second round-trip) or SaveImage for masks, and the builder refuses to emit a graph that would run green and return nothing.

Provenance (new in 3.1)

ComfyUI embeds the exact API-format graph in every output PNG. comfy-headless reads it back — pure stdlib, no Pillow — and can re-run it verbatim:

from comfy_headless import read_workflow_metadata, extract_prompt_graph

record = read_workflow_metadata("output.png")
print(record.prompt is not None)     # the machine-runnable graph
print(record.extra)                  # your custom keys land here

graph = extract_prompt_graph("output.png")   # raises with a hint if scrubbed
result = client.rerun_from_png("output.png") # re-POSTs it verbatim

Write custom provenance without any custom node — anything in extra_pnginfo becomes a PNG text chunk in the outputs:

client.queue_prompt(workflow, extra_pnginfo={"myapp:run_id": "r-2026-077"})

Known limits, documented rather than hidden: WebP/JPEG carry the same data in EXIF (a different reader path, not implemented); video outputs don't embed the graph; hardened deployments may strip unknown keys; and GUI→API conversion has no server route — use ComfyUI's "Workflow → Export (API)".

Configuration

Environment variables use the COMFY_HEADLESS_ prefix with __ section delimiters:

Variable	Default
COMFY_HEADLESS_COMFYUI__URL	http://localhost:8188
COMFY_HEADLESS_OLLAMA__URL	http://localhost:11434
COMFY_HEADLESS_OLLAMA__MODEL	qwen2.5:7b
COMFY_HEADLESS_COMFYUI__TIMEOUT_READ	read timeout, seconds
COMFY_HEADLESS_LOGGING__LEVEL	log level

Or pass the URL directly: ComfyClient("http://192.168.1.50:8188").

Web UI

comfy-headless                 # launching the UI is the default action

Flag	Meaning
--port / -p	UI port (default 7861)
--share	Public Gradio share link
--url	ComfyUI server URL
--version / -v	Print version
--check	Feature availability
--diagnose	Full diagnostics

Six tabs: Image, Video, Queue & History, Workflows, Models, Settings. Theme is Ocean Mist — soft teal accents on warm neutral backgrounds.

Programmatically (requires [ui]):

from comfy_headless import launch
launch(port=7861, share=False)

Progress

Blocking calls take an on_progress callback:

client.generate_image("a fox", on_progress=lambda pct, msg: print(f"{pct:.0%} {msg}"))

For real-time updates over WebSocket (requires [websocket]):

import asyncio
from comfy_headless import ComfyWSClient

async def main():
    async with ComfyWSClient() as ws:
        prompt_id = await ws.queue_prompt(workflow)
        return await ws.wait_for_completion(prompt_id)

asyncio.run(main())

Errors

Every exception carries a structured code, message and hint:

from comfy_headless import (
    ComfyHeadlessError,       # base
    ComfyUIConnectionError,   # cannot reach ComfyUI
    ComfyUIOfflineError,      # ComfyUI not responding
    GenerationTimeoutError,
    GenerationFailedError,
    ValidationError,
    UploadError,              # new in 3.0
    MissingNodePackError,     # new in 3.0
    GraphAddressError,        # new in 3.1 — bad dotted field / inactive combo branch
)

try:
    client.generate_image("test")
except ComfyUIOfflineError:
    print("Start ComfyUI first")

How it works

your call ─→ build API-format graph ─→ POST /prompt ─→ poll /history ─→ GET /view
                     │
                     └─ validated against GET /object_info

The library talks to seven ComfyUI routes — /system_stats, /object_info, /queue, /history, /prompt, /interrupt, /view — plus /upload/image and /upload/mask for binary input (audio uploads ride /upload/image too; the server has no audio-specific route). All six profiles fit inside that surface: v3.1 added meshes, music, captions and provenance without adding a single route.

/object_info is the authority on what a given server can run. It is a live endpoint, not a versioned artifact: there is no core-node registry to pin against. So the library validates emitted graphs against the target server's actual catalog rather than assuming a fixed node set. check_workflow_dependencies() is that check, and it is load-bearing infrastructure rather than a convenience.

Useful escape hatches when you want the graph itself:

workflow = client.build_txt2img_workflow("a fox")   # the raw API-format dict
prompt_id = client.queue_prompt(workflow)           # submit it yourself
client.wait_for_completion(prompt_id)

Docs

Full handbook: mcp-tool-shop-org.github.io/comfy-headless — getting started, usage, the six profiles, video models, configuration, API reference, architecture.

In-repo knowledge base for LLMs and contributors: kb/ — a machine-readable index.json over per-profile fact pages, runnable reference graphs (kb/workflows/*.json, generated from the builders themselves so they cannot drift), and node provenance (kb/nodes.json). python scripts/gen_kb.py regenerates it; the test suite fails if code and KB disagree.

Security & data scope

• Data touched: connects to a local or remote ComfyUI instance over HTTP/WebSocket. Sends workflow JSON and uploaded images, receives generated media. Optionally connects to a local Ollama for prompt intelligence. Writes output to temp directories with automatic cleanup.
• Data NOT touched: no telemetry, no analytics, no external APIs beyond the ComfyUI and Ollama endpoints you configure. Secrets are masked in all log output via SecretValue.
• Permissions required: network access to your ComfyUI server and optional Ollama server; file write for output and temp directories.
• Uploads: upload_image rejects subfolder traversal attempts. Uploaded files land in ComfyUI's input directory on whichever server you point at — treat that server as trusted.

See SECURITY.md for vulnerability reporting.

Scorecard

Category	Score
A. Security	10/10
B. Error Handling	10/10
C. Operator Docs	10/10
D. Shipping Hygiene	10/10
E. Identity (soft)	10/10
Overall	50/50

Assessed with @mcptoolshop/shipcheck

Related

Part of MCP Tool Shop — open-source ML tooling for local hardware.

Contributing

Issues and pull requests welcome — see CONTRIBUTING.md. Useful areas: additional model families, workflow templates, docs, bug fixes.

If you add a node type, verify it exists in the live ComfyUI catalog first, and declare its pack if it isn't core. That rule is why this release exists.

License

MIT — see LICENSE.

---

Built by MCP Tool Shop