comfyanonymous/ComfyUI
Fun facts
Light trivia about the codebase. Not load-bearing.
"comfyanonymous"
The project's namesake author goes by comfyanonymous everywhere — the GitHub handle, the git author name, the repo URL. As of HEAD they have authored 3,125 of 5,110 commits (61%) — solo project growth that turned into a full ecosystem with the Comfy-Org organization layered around it (frontend, desktop, manager, cloud).
The first commit
2023-01-03 220afe33 Initial commit.By 2023-01-22 ComfyUI already had:
- A
LoadImagenode for img2img - The
IS_CHANGEDmechanism for incremental re-execution - An execution scheduler that "tries to execute the outputs that take the less steps first"
- Single-line string properties on nodes
In other words, the graph executor with caching — still the heart of the codebase — existed within three weeks of the first commit.
The biggest file in the repo is a stub
comfy_api_nodes/apis/__init__.py weighs in at 6,155 lines. It is a single file of generated Pydantic models for every paid-API integration in comfy_api_nodes/. The comfy_api/generate_api_stubs.py script regenerates these stubs and the update-api-stubs.yml workflow runs them.
The runner-up is more representative of the codebase: comfy_api_nodes/nodes_kling.py at 3,327 lines — one file for one paid video provider. Several other paid-API node modules (Wan, ByteDance, Vidu) clear 1,500 lines too.
ComfyUI as a model release seismograph
comfy/supported_models.py is 1,884 lines of ModelXXX classes — one entry per architecture the engine has ever shipped support for. Walking comfy/text_encoders/ and comfy/ldm/ is essentially a year-by-year list of which generative models mattered to the team:
- Image: SD1.x, SD2.x, SDXL, Stable Cascade, Pixart, AuraFlow, HunyuanDiT, SD3/3.5, Flux (and Flux 2, Kontext, Krea), Lumina 2, HiDream, Qwen Image, Hunyuan Image 2.1, Z Image, Chroma Radiance, Omnigen 2.
- Video: Stable Video Diffusion, Mochi (Genmo), LTX-Video (Lightricks), Hunyuan Video, Wan 2.1/2.2, Hunyuan Video 1.5, Kandinsky 5, Cosmos.
- Audio: Stable Audio, ACE Step.
- 3D: Hunyuan3D 2.0/2.1.
If you ever forget when something dropped, git log --diff-filter=A -- comfy/ldm/<model>/ will tell you.
A "wiki bot" is the most active non-comfyanonymous author
Of the 5,110 commits, the second-most-prolific named author after comfyanonymous is:
| Author | Commits |
|---|---|
comfyanonymous |
3,125 |
| Alexander Piskun | 256 |
| pythongosssss | 231 |
| Chenlei Hu | 118 |
| rattus | 105 |
ComfyUI Wiki |
98 |
That last one is a bot that mirrors the GitHub Wiki pages back into the repo. It is one of only three "bot" authors in the entire history (118 commits combined out of 5,110 — about 2.3%).
The hook_breaker_ac10a0 module
There is a top-level file called hook_breaker_ac10a0.py — 17 lines — whose entire job is to save and restore monkey-patched function references that custom nodes might have installed. The cryptic suffix is unique enough to make collisions with custom-node names impossible. It runs around prompt execution in main.py so that one badly-behaved custom node can't hijack the runtime forever.
Telemetry-off, on principle
main.py sets two environment variables before importing anything that might phone home:
os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
os.environ['DO_NOT_TRACK'] = '1'The README also brags: "Works fully offline: core will never download anything unless you want to." The frontend manager is the one place that opts back into network access — and only to fetch frontend releases.
The 16,384 pixel limit
nodes.py defines MAX_RESOLUTION = 16384. Every built-in node that takes a width or height clamps to this. Yes, it is exactly 2¹⁴.
Built by Factory AutoWiki from public repository content. It is a generated preview for codebase exploration, not source-maintained documentation.