A dedicated server runs ComfyUI for image generation workflows — completely separate from the language model servers, on an NVIDIA GPU.
AI3 is equipped with an NVIDIA RTX 5060 Ti with 16 GB VRAM — chosen specifically for CUDA-based image generation workloads. The system runs an Intel i5-14400F with 16 GB DDR5 RAM.
The NVIDIA GPU was a deliberate choice alongside the AMD Radeon cards on AI1 and AI2: image generation workflows benefit strongly from CUDA and the broader model ecosystem built around it.
ComfyUI is the interface of choice for its node-based workflow editor. Pipelines can be constructed visually — chaining samplers, conditioning, upscalers, and ControlNet in a reusable graph format.
ComfyUI runs as a service on Ubuntu 24.04 and is accessible on the internal network. No external access is exposed.
Image generation and LLM inference have different hardware profiles. LLMs benefit from large VRAM (32 GB per GPU on AI1 and AI2), while image generation on modern models runs efficiently on 16 GB with CUDA. Running both workloads on the same machine would mean GPU context switching and slower responses for everyone.
Keeping AI3 dedicated means image generation never competes with ongoing conversations on AI1 or AI2. The three servers run independently, each optimised for its workload.