Get your key from BytePlus Console. Required for SeedDream 5.0 Lite ($0.035/image).

Get your key from fal.ai Dashboard. Required for Chatterbox voice changer (from $0.015/min) and Whisper captions.

The only thing you ever enter. From this key PatchWork finds your pods, starts one when you need a GPU, works out its address itself, and terminates it when you are done — you never paste a pod URL, because a pod URL dies with its pod. Get a key from RunPod → Settings. Stored on this device only and never synced. Pods bill by the second and terminate themselves when idle, when this tab stops responding, or after 2 hours.

Set up finds the volume on your RunPod account or creates one (20 GB, $1.40/month, billed whether or not a pod is running — the only charge here that does not stop with the pod). It never makes a second one: it matches by name first, so pressing it twice adopts what is already there. A RunPod network volume holding the Python environment as a single archive, restored in ~24 s instead of the ~81 s it takes to pip-install from scratch. Measured live: boot to ready fell from 248 s to 118 s. Both fields or neither — a volume can only be mounted from the datacenter that holds it, so an ID without a region is treated as no cache. Note that setting one pins launches to that datacenter, giving up the automatic retry into the second one when capacity runs out. Leave blank to always build fresh. The 63 GB of model weights are never cached here — they download faster than this volume can serve them.

Comes pre-filled with the two H3 adapters worth having — Realism (people) and Anime (sketch). Clear the box to boot pods with none. One per line: repo | filename | name | trigger words. Only the first two are required. Each is downloaded from Hugging Face when a pod boots, then offered on any H3 node under STYLE LORAS, where they stack on top of the Turbo switch. Trigger words are added to the prompt automatically when the LoRA is selected — a LoRA does nothing until its token appears. LoRAs are architecture-specific: one trained for Wan or Flux will not load on MiniMax H3. Pods pick up library changes at boot, so a running pod keeps whatever it started with.

How ComfyUI computes attention. Sol is the sparse sol-attn patch (the tested default); None is plain PyTorch — no patch, no flag — the fastest at full quality on this card. Sol and None share a pod, so switching between them is free; only Sage is a launch flag, so switching to or from Sage means a new pod. Sage is not recommended yet. ComfyUI issue #15263 reports it turning MiniMax H3 into pure noise — video and audio both — on sm_120, which is the RTX PRO 6000 we run, above roughly 167k tokens. It behaves as a cliff rather than a slope, so a 5s clip can look perfect and a 15s clip be unusable at identical settings. Reported fixed upstream, but we clone ComfyUI unpinned and have never measured it here. Leave on Sol unless you are deliberately testing.

Paste the address of a running ComfyUI server to enable the MiniMax H3 models. On RunPod this is https://<podId>-8188.proxy.runpod.net. The server must be started with --listen 0.0.0.0 --enable-cors-header. This URL is stored on this device only and is never synced — it stops working as soon as the pod is terminated.

Every pod generation records a row — mode, prompt, preset, step count, frame count, seed, and both the server's own execution time and the wall time you waited. Exports as CSV for working out what a clip actually costs.

Get your key from Resemble AI Dashboard. Alternative voice changer via Speech-to-Speech (~$0.006/sec).

Get your key from ElevenLabs → Profile → API key. Speech-to-Speech voice conversion (~$0.17/min on Starter plan).