Deploying ComfyUI on Kubernetes for Game Art Production

[ShawnDen-coder]

Hello 👋
I'm from the game industry. Our art team uses ComfyUI daily for concept art generation, asset creation, and style transfer in our production pipeline.

What we want to do
We plan to deploy ComfyUI on Kubernetes (K8s) to leverage:

Dynamic GPU Resource Allocation — Scale worker pods up/down based on workload demand, avoiding idle GPU waste.

API Exposure — Expose ComfyUI's /prompt API via K8s Ingress/Service, allowing our internal tools and automated pipelines to submit workflows programmatically.

Multi-team Isolation — Different art teams (character, environment, UI) get isolated namespaces with dedicated resources.

Questions
Has anyone deployed ComfyUI on K8s in production? Any tips or pitfalls?
Best practice for sharing model files (checkpoints/LoRA) across multiple pods?
How do you handle cold start / model loading time when scaling up new pods?
Contribution
I'm very willing to open a branch and start working on this myself. That said, if anyone in the community has experience or interest in this area, I'd love to collaborate — any help or guidance would be greatly appreciated!

Thanks! Happy to share our progress along the way. 🎮

[xiaoshutu2]

Hello, may I ask if you have encountered a situation where the Anmia large model is incompatible with CLIP text during deployment? If so, how did you handle it?