KinkzAI Control

Standard model Premium model Box image (decides what's already installed)

Nothing about the machine changes this — it's chosen when renting. vllm/vllm-openai:latest has vLLM but no nginx (setup installs it). Vast's own vllm template bundles more but is slower to boot.

Voices zip link Voice model Turbo — 350M, what boxes run today Nano — 110M, faster, slightly flatter

Same server and the same voices either way — only the checkpoint changes. Applies when a box is set up; change it and re-run setup to switch an existing box.

Also install voice on model boxes

Off is right once you have a voice box. Voice on every model box reserves memory that comes straight out of KV cache — so it costs concurrent conversations on every card, not just setup time.

Generation — version (unused, ComfyUI tracks master)

A git tag, pinned for the same reason the vLLM image is: a version that works keeps working. Bump it deliberately, not by accident.

Generation — model manifest

Fetched into ComfyUI's models folder at setup, so a fresh box arrives stocked — boxes are disposable, the manifest is what persists. Common subfolders: Stable-Diffusion (checkpoints), diffusion_models (video/unet files), Lora, VAE, clip, upscale_models. Gated downloads need HF_TOKEN / CIVITAI_TOKEN in the control server's .env — tokens stay there, never in this file.

Generation — box image

The box builds ComfyUI's venv and picks its own torch, so the image just needs to be a pinned, CUDA-ready base that doesn't move.

Takes effect on the next box you set up. Boxes already running keep what they were built with.