Skip to content

comfyui

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

comfyui

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
comfyui-deployment.yaml
comfyui-deployment.yaml
 
 
comfyui-service.yaml
comfyui-service.yaml
 
 
namespace.yaml
namespace.yaml
 
 

README.md

ComfyUI Workload

ComfyUI — a node-based UI for diffusion models and other generative AI workflows.

Deployment

The manifests create:

Docker Image Selection

Initially tried ghcr.io/ai-dock/comfyui:pytorch-2.5.1-cu124-runtime but ran into issues. Switched to mmartial/comfyui-nvidia-docker:ubuntu24_cuda12.8-latest which provides:

  • Ubuntu 24 + CUDA 12.8 base
  • FORCE_CHOWN=yes env var for fixing volume ownership on startup
  • USE_UV=true for faster Python package installs
  • Built-in venv at /comfy/mnt/venv

Deploying

kubectl apply -f namespace.yaml
kubectl apply -f comfyui-deployment.yaml
kubectl apply -f comfyui-service.yaml

Access at http://<node-ip>:30188

Use Cases

ComfyUI is used for three different projects in this lab:

Project What It Does Custom Nodes
Dataset Creation Qwen Image Edit + Florence2 captioning pipeline ComfyUI-GGUF, ComfyUI-Florence2, ComfyUI-KJNodes, comfyui_controlnet_aux, rgthree-comfy
Music Generation ACE-Step 1.5 audio generation ComfyUI_ACE-Step
LoRA Inference Generate images using trained LoRA adapters (built-in LoraLoader node)

Custom Node Installation

All custom nodes were installed via git clone into the pod's /workspace/custom_nodes/ directory, then dependencies installed with the venv pip.

Batch install (for Qwen Image Edit pipeline)

POD=$(kubectl -n comfyui get pod -l app=comfyui -o jsonpath='{.items[0].metadata.name}')

kubectl exec -n comfyui $POD -- bash -c '
  cd /workspace/custom_nodes &&
  git clone https://github.com/city96/ComfyUI-GGUF.git &&
  git clone https://github.com/rgthree/rgthree-comfy.git &&
  git clone https://github.com/kijai/ComfyUI-KJNodes.git &&
  git clone https://github.com/Fannovel16/comfyui_controlnet_aux.git
'

Florence2 (for auto-captioning)

kubectl exec -n comfyui $POD -- bash -c '
  cd /workspace/custom_nodes &&
  git clone https://github.com/kijai/ComfyUI-Florence2.git &&
  cd ComfyUI-Florence2 &&
  /comfy/mnt/venv/bin/pip install -r requirements.txt
'

Dependencies: timm, peft, accelerate. The Florence2 model itself (microsoft/Florence-2-base) downloads automatically on first workflow run.

ACE-Step (for music generation)

kubectl exec -n comfyui $POD -- bash -c '
  cd /workspace/custom_nodes &&
  git clone https://github.com/billwuhao/ComfyUI_ACE-Step.git &&
  cd ComfyUI_ACE-Step &&
  /comfy/mnt/venv/bin/pip install -r requirements.txt
'

ComfyUI Manager

kubectl exec -n comfyui $POD -- bash -c '
  cd /workspace/custom_nodes &&
  git clone https://github.com/ltdrdata/ComfyUI-Manager.git
'

After installing nodes — restart the pod

kubectl -n comfyui rollout restart deployment comfyui

Custom Nodes Summary

Node Repo Purpose Installed For
ComfyUI-GGUF city96 Load GGUF-quantized models Qwen Image Edit
ComfyUI-Florence2 kijai Florence2 image captioning Dataset creation
ComfyUI-KJNodes kijai LoadImagesFromFolder, JoinStrings, ImageResize, StringConstant Dataset creation
comfyui_controlnet_aux Fannovel16 ControlNet preprocessors (DWPose, depth) Qwen Image Edit
rgthree-comfy rgthree SetNode/GetNode for clean wiring Qwen Image Edit
ComfyUI_ACE-Step billwuhao ACE-Step music generation Music generation
ComfyUI-Manager ltdrdata Package manager UI General
comfyui-videohelpersuite (pre-installed) Video processing Came with Docker image

Model Downloads

All models were downloaded manually via kubectl exec + wget into the pod. Here are the exact commands.

Qwen Image Edit Models

# GGUF model (14 GB) → /workspace/models/unet/gguf/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/unet/gguf &&
  wget -q -c -O /workspace/models/unet/gguf/Qwen-Image-Edit-2509-Q5_0.gguf \
    "https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/resolve/main/Qwen-Image-Edit-2509-Q5_0.gguf"
'

# CLIP encoder (8.7 GB) → /workspace/models/clip/qwen/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/clip/qwen &&
  wget -q -c -O /workspace/models/clip/qwen/qwen_2.5_vl_7b_fp8_scaled.safetensors \
    "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors"
'

# VAE (243 MB) → /workspace/models/vae/qwen/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/vae/qwen &&
  wget -q -c -O /workspace/models/vae/qwen/qwen_image_vae.safetensors \
    "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors"
'

# Lightning LoRA for 4-step generation (811 MB) → /workspace/models/loras/qwen/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/loras/qwen &&
  wget -q -c -O /workspace/models/loras/qwen/Qwen-Image-Lightning-4steps-V2.0.safetensors \
    "https://huggingface.co/lightx2v/Qwen-Image-Lightning/resolve/main/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors"
'

ACE-Step Models

# ACE-Step 1.5 Turbo AIO checkpoint (9.4 GB) → /workspace/models/checkpoints/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/checkpoints &&
  wget -q -c -O /workspace/models/checkpoints/ace_step_1.5_turbo_aio.safetensors \
    "https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/resolve/main/checkpoints/ace_step_1.5_turbo_aio.safetensors"
'

# ACE-Step v1 checkpoint (7.2 GB) → /workspace/models/checkpoints/
kubectl exec -n comfyui $POD -- bash -c '
  wget -q -c -O /workspace/models/checkpoints/ace_step_v1_3.5b.safetensors \
    "https://huggingface.co/Comfy-Org/ACE-Step_ComfyUI_repackaged/resolve/main/all_in_one/ace_step_v1_3.5b.safetensors"
'

Upscale Model

# 4x-UltraSharp (67 MB) → /workspace/models/upscale_models/
kubectl exec -n comfyui $POD -- bash -c '
  mkdir -p /workspace/models/upscale_models &&
  wget -q -c -O /workspace/models/upscale_models/4x-UltraSharp.pth \
    "https://huggingface.co/madriss/chkpts/resolve/main/ComfyUI/models/upscale_models/4x-UltraSharp.pth"
'

Model Directory Layout

/workspace/models/
├── checkpoints/
│   ├── ace_step_1.5_turbo_aio.safetensors    # 9.4 GB — ACE-Step 1.5 Turbo
│   └── ace_step_v1_3.5b.safetensors          # 7.2 GB — ACE-Step v1
├── clip/qwen/
│   └── qwen_2.5_vl_7b_fp8_scaled.safetensors # 8.7 GB — Qwen CLIP encoder
├── loras/qwen/
│   └── Qwen-Image-Lightning-4steps-V2.0.safetensors  # 811 MB — Lightning LoRA
├── unet/gguf/
│   └── Qwen-Image-Edit-2509-Q5_0.gguf        # 14 GB — Qwen Image Edit GGUF
├── upscale_models/
│   └── 4x-UltraSharp.pth                     # 67 MB — 4x upscaler
└── vae/qwen/
    └── qwen_image_vae.safetensors            # 243 MB — Qwen VAE

Total model storage: ~40 GB

Permission Gotcha

The mmartial/comfyui-nvidia-docker image runs as a non-root user (comfytoo). When git-cloning custom nodes or downloading models into hostPath volumes, you may hit permission errors. Two fixes:

  1. Set FORCE_CHOWN=yes in the deployment env (already in our manifest) — fixes ownership on pod startup
  2. Manual fix inside the pod if needed: 
    kubectl exec -n comfyui $POD -- sudo chown -R comfytoo:comfytoo /workspace/custom_nodes
kubectl exec -n comfyui $POD -- sudo chown -R comfytoo:comfytoo /comfy/mnt/venv

GPU Assignment

Unlike AI-Toolkit, ComfyUI does not use UUID pinning. It requests nvidia.com/gpu: "1" and gets whichever GPU the Kubernetes scheduler assigns. This is fine for inference workloads that fit in 8-10 GB VRAM.

Storage

Uses hostPath volumes:

  • /data/comfyui → /workspace (models, outputs, custom nodes)
  • /data/comfyui/run → /comfy/mnt (runtime data, venv)
  • dshm emptyDir → /dev/shm (8 GiB shared memory)