ComfyUI Text-to-Image/Video Workload¶

This Helm Chart deploys a ComfyUI web app for text-to-image/video generation. ComfyUI is a powerful node-based interface for stable diffusion that provides advanced workflows for AI image and video generation.

Features¶

Pre-configured ComfyUI Environment: Automatically installs and configures ComfyUI with ROCm support
Model Management: Support for downloading models from Hugging Face or MinIO/S3 storage
ComfyUI Manager: Includes ComfyUI Manager for easy extension management

Configuration Parameters¶

You can configure the following parameters in the values.yaml file or override them via the command line:

Parameter	Description	Default
`image`	Container image repository and tag	`rocm/pytorch:rocm7.1.1_ubuntu24.04_py3.12_pytorch_release_2.8.0`
`imagePullSecrets`	List of Kubernetes secrets for pulling images from private registries	`[]`
`gpus`	Number of GPUs to allocate	`1`
`model`	Hugging Face model path (e.g., `Comfy-Org/flux1-dev`). Set to `""` to start with no checkpoint	`Comfy-Org/stable-diffusion-v1-5-archive`
`tag`	Identifies one model's binaries (tagsafetensors) to download. Without it the whole repository is downloaded	`v1-5-pruned-emaonly-fp16`
`storage.ephemeral.quantity`	Ephemeral storage size	`200Gi`
`kaiwo.enabled`	Enable Kaiwo operator management	`false`
## Using Private Container Registries

If you need to pull images from a private registry, set the imagePullSecrets field in your values.yaml or via the command line. This should be a list of Kubernetes secret names that provide credentials for your registry.

Example in values.yaml:

imagePullSecrets:
  - my-registry-secret

Or via the command line:

helm template . --set imagePullSecrets={my-registry-secret} | kubectl apply -f -

The deployment will include these secrets in the pod spec, allowing Kubernetes to authenticate to your private registry.

Environment Variables¶

The following environment variables are configured for MinIO/S3 integration:

Variable	Description	Default
`BUCKET_STORAGE_HOST`	MinIO/S3 endpoint URL	`http://minio.minio-tenant-default.svc.cluster.local:80`
`BUCKET_STORAGE_ACCESS_KEY`	MinIO/S3 access key (from secret)	From `minio-credentials` secret
`BUCKET_STORAGE_SECRET_KEY`	MinIO/S3 secret key (from secret)	From `minio-credentials` secret
`PIP_DEPS`	Additional Python packages to install via pip (space or newline separated URLs/packages)	`""`
`COMFYUI_PATH`	ComfyUI installation path	`/workload/ComfyUI`
`MODEL_BIN_URL`	Direct URL to download an additional model checkpoint (optional)	Not set

Model Configuration¶

The default deployment pre-loads v1-5-pruned-emaonly-fp16.safetensors (2 GiB), which is the checkpoint ComfyUI's stock workflow selects by name, so the workspace can generate an image as soon as it opens. The workspace reports itself ready only once that checkpoint is on disk.

tag must identify a single model's file. It is matched as *tag*safetensors both to choose what to download and to decide the workspace is ready, so a fragment shared by several models (fp8, say) would let a checkpoint left by a previously configured model pass for the current one.

Other ways to get a model:

A different checkpoint at deploy time, by setting model and tag, or by using one of the overrides in overrides/models/. Larger models take proportionally longer before the workspace becomes ready, and the stock workflow will need its checkpoint re-selected.
At runtime from the UI, using the ComfyUI-Manager model manager. Set model: "" to skip the pre-load entirely, in which case the workspace becomes ready as soon as the server answers and starts with an empty checkpoint list.

Using Hugging Face Models¶

Configure models from Hugging Face by setting the model parameter:

# Example: FLUX.1-dev model
model: "Comfy-Org/flux1-dev"
tag: "flux1-dev-fp8"

The example above appears in ComfyUI as flux1-dev-fp8.safetensors. Note the tag is the full flux1-dev-fp8 rather than fp8, which Comfy-Org/flux1-schnell also matches.

Using S3/MinIO Models¶

For models stored in S3/MinIO, use the s3:// prefix:

model: "s3://models/Comfy-Org/flux1-dev"

Using Direct Download URLs¶

For direct model downloads, use the MODEL_BIN_URL environment variable:

env_vars:
  MODEL_BIN_URL: "https://huggingface.co/Comfy-Org/Lumina_Image_2.0_Repackaged/resolve/main/all_in_one/lumina_2.safetensors"

Pre-configured Model Overrides¶

The workload includes several pre-configured model overrides in the overrides/models/ directory:

Deploying the Workload¶

Basic Deployment¶

To deploy the service with default settings, run the following command within the helm folder:

helm template . | kubectl apply -f -

Deployment with Model Override¶

To deploy with a specific model configuration:

helm template flux . -f overrides/models/comfy-org_flux1-dev-fp8.yaml | kubectl apply -f -

Custom Deployment¶

To deploy with custom parameters. Set tag whenever you change model, so that it identifies a file the new repository actually contains:

helm template flux . --set model="Comfy-Org/flux1-dev" --set tag="flux1-dev-fp8" | kubectl apply -f -

Accessing the Workload¶

Verify Deployment¶

Check the deployment and service status:

kubectl get deployment
kubectl get service

Port Forwarding¶

To access the service locally on port 8188, forward the port of the service/deployment:

kubectl port-forward services/dev-text2image-comfyui 8188:80

Then open a web-browser and navigate to http://localhost:8188 to access ComfyUI.

Accessing the Workload via URL¶

To access the workload through a URL, you can enable either an Ingress or HTTPRoute in the values.yaml file. The following parameters are available:

Parameter	Description	Default
`ingress.enabled`	Enable Ingress resource	`false`
`http_route.enabled`	Enable HTTPRoute resource	`false`
`http_route.parentRefs`	List of gateway parent references (`group`, `name`, `namespace`), takes precedence over gateway namespace keys	`[]`
`http_route.gatewayNamespace`	Single-gateway fallback used when `parentRefs` is empty	`envoy-gateway-system`
`http_route.gateway_namespace`	Deprecated alias for `gatewayNamespace`	`""`

Example dual-gateway configuration:

http_route:
  enabled: true
  parentRefs:
    - group: gateway.networking.k8s.io
      name: https
      namespace: envoy-gateway-system
    - group: gateway.networking.k8s.io
      name: https
      namespace: kgateway-system

See the corresponding template files in the templates/ directory. For more details on configuring Ingress or HTTPRoute, refer to the Ingress documentation and HTTPRoute documentation, or documentation of the particular gateway implementation you may use, like KGateway. Check with your cluster administrator for the correct configuration for your environment.

Health Checks and Monitoring¶

The workload includes comprehensive health monitoring:

Startup Probe: Allows up to 10 minutes for ComfyUI to start (checks /queue endpoint)
Liveness Probe: Monitors if ComfyUI is running properly
Readiness Probe: Asks ComfyUI which checkpoints it can see (/models/checkpoints). When model is set, it requires that model's checkpoint, so the workload only receives traffic once it is usable. With no model, the server answering is enough

A configured checkpoint is downloaded in the background so that a large model cannot delay the server bind past the startup probe budget. ComfyUI therefore serves /queue before the model is on disk, and the pod only becomes Ready once the checkpoint appears. A download that fails or stalls restarts the container, which resumes the transfer, rather than leaving a modelless workload running.