Infinity Engine i C. infinity

Set up a GPU (spinning up other devices)

The engine does all its thinking on your own machine. The only step that needs a real GPU is the render: turning a job's shots into images or video through ComfyUI. This page is the concrete "spin up another device" story, from renting a box to frames landing back on your laptop.

You never install ComfyUI on your laptop. You rent a GPU box by the hour, run ComfyUI on it, and point the engine at it. When the batch is done, you stop the box and stop paying.

Prefer buttons? Use the control panel

Everything on this page can be done without the terminal. Double-click Infinity Engine.cmd in the repo folder (or run `python -m engine gui`) and a control panel opens in your browser: pick a song, analyse it, make a job, render offline or on a pod, advance a stage, and see the wiring, all as buttons. The panel is local only; it never gets published. The commands below are the same actions, for when you are on the GPU box itself or prefer typing.

The one touchpoint

Everything ComfyUI-related goes through a single URL. Check it any time:

engine doctor

That prints where the engine expects ComfyUI (server in catalog/comfy.yaml, default http://127.0.0.1:8188), whether it is reachable right now, and which checkpoint each recipe expects. When reachable says YES, renders will run. Until then, `engine work --offline` still writes the exact ComfyUI graphs it would submit, so you can build and inspect the whole pipeline with no GPU at all.

Rent a box

Any Ubuntu + NVIDIA host works. Good picks (prices AUD, re-check before a big batch):

best for cheap draft batches.

~A$4.30/hr; has a Sydney region and network volumes that keep weights warm between sessions.

all (a different path; this page covers the rent-a-box path).

Pick a GPU with enough VRAM for the recipe tier you want (the vram_gb in providers.yaml): a 4090 (24GB) covers draft and standard image work.

Bootstrap it

Copy tools/pod_bootstrap.sh onto the box and run it once:

bash pod_bootstrap.sh

It installs ComfyUI + the Manager and starts the server on 127.0.0.1:8188. Then you download the weights your recipes name (it prints where); engine doctor lists exactly which ones. Leave ComfyUI running.

Connect: two ways

A. SSH tunnel (simplest, recommended to start). From your laptop, open a tunnel so the pod's ComfyUI looks local:

ssh -L 8188:localhost:8188 root@POD_ADDRESS

Leave that open. Now the engine on your laptop can drive the pod:

engine doctor                 # reachable should now say YES
engine make SLUG panels       # assemble a job
engine work                   # renders on the pod, frames land locally

The job folder never leaves your machine; only the graph travels over the tunnel as an HTTP call, and the images come straight back into the job's results/. This is the whole loop, tethered to the pod only while it runs.

B. Point straight at the box. If you would rather not tunnel, expose the pod's port and pass the address:

engine work --server http://POD_ADDRESS:8188

Same result. Use a tunnel or a firewall rule so the ComfyUI port is not open to the world.

C. Run the engine on the pod (for long unattended batches). Copy the repo (or just the queued job folders) to the box, run engine work there against its local ComfyUI, then copy results/ back. Best when you do not want your laptop tethered for hours.

The full loop, once a box is up

engine make SLUG panels        # 1-5: data, model, compute, direction -> job folder
engine work                    #  6 : render on the GPU (or --offline to just emit graphs)
#          review results/ and the manifest
engine advance SLUG panels     #  7 : move the song to the next stage

Stop the box when the batch is done. Nothing on your side depends on it staying up: the vault, the jobs and their results are all local.

Which model renders

You never edit code to swap a model. Each recipe in catalog/comfy.yaml names a ckpt_name (the exact file under ComfyUI/models/checkpoints/) and its sampler settings, keyed by category and tier. Change the model = change that line, put the weights on the pod, engine doctor to confirm. Qwen-Image and FLUX.2 use different loader nodes than the classic graph; those plug in as new graph builders in engine/comfy.py, still with no change to the render loop.