ComfyUI

The image renderer behind darkroom, atlas bakes and proof's evaluation runs. Darkroom needs a websocket output node and two of its own custom nodes, which a stock install lacks.

Used by

Package	What for	Reads
argus-darkroom (opens in new tab)	Required. Every generation. Needs `SaveImageWebsocket`, plus darkroom's custom nodes for image-to-image and inpainting.	`ARGUS_DARKROOM_COMFYUI_URL`, `ARGUS_DARKROOM_WORKFLOW`
argus-atlas (opens in new tab)	`bake`: the map restyle (off unless `ATLAS_ALLOW_BAKE=1`).	`COMFYUI_BASE_URL`, `--comfyui-url`, `ATLAS_BACKEND=comfyui`, `ATLAS_WORKFLOW`, `ATLAS_ALLOW_BAKE`
argus-proof (opens in new tab)	Renders the evaluation prompts with the trained LoRA (the default backend).	`PROOF_BACKEND=comfyui`, `COMFYUI_BASE_URL`, `PROOF_WORKFLOW_TEMPLATE`
argus-forge (opens in new tab)	Trained LoRAs land in `<export>/forge/kohya/output/`. Point that volume or `FORGE_PATH_MAP` at ComfyUI's `models/loras/`.	`FORGE_PATH_MAP`
argus-mint (opens in new tab)	Renders through argus-darkroom, so indirectly.	`ARGUS_MINT_DARKROOM_URL`

Official guides

ComfyUI on GitHub (opens in new tab)
System requirements (opens in new tab)
Manual install (opens in new tab): also the `models/` folder layout
Custom nodes (opens in new tab)

Minimal setup

Install ComfyUI and start it (python main.py --listen). Use --listen so containers can reach it. It listens on port 8188.
Put your checkpoints and LoRAs in its models/ tree (see Checkpoints and LoRAs).

Point each client at it. The variable names and defaults differ:

Package	Variable	Default
darkroom	`ARGUS_DARKROOM_COMFYUI_URL`	`http://localhost:8188`
atlas	`COMFYUI_BASE_URL`	`http://host.docker.internal:8188`, even when run natively
proof	`COMFYUI_BASE_URL`	`http://127.0.0.1:8188`
Studio compose	`COMFYUI_BASE_URL` (passed to all three)	`http://host.docker.internal:8188`

For a ComfyUI on another machine, use http://<GPU_HOST>:8188.

Darkroom's nodes

Darkroom receives images over ComfyUI's websocket, so a stock install needs two additions.

SaveImageWebsocket ships with ComfyUI as an example node (custom_nodes/websocket_image_save.py). Check that it's loaded:
```
curl http://localhost:8188/object_info/SaveImageWebsocket
```
DarkroomLoadImageBase64 and DarkroomLoadMaskBase64 are needed only for image-to-image and inpainting. They live in darkroom's repo, outside the Python package. Copy the file in and restart ComfyUI:
```
cp argus-darkroom/comfyui/darkroom_load_image_base64.py <ComfyUI>/custom_nodes/
curl http://localhost:8188/object_info/DarkroomLoadMaskBase64
```
If you copied an older version before the mask node existed, copy it again.

Atlas's workflow

Atlas's starter workflow template names an SDXL depth ControlNet. Swap in one your ComfyUI actually has. See the ComfyUI workflow template (opens in new tab).