kazeizumi H3 runtime router
When the user asks to generate H3 video without specifying the runtime, ask one concise question: cloud or local? If the current request already names one, use it and do not ask again.
Build the generation contract
Before any real submission, identify and confirm the following for the current run:
- backend and exact workflow name or file;
- reference files and real upload order;
- prompt and parameter overrides;
- job count and duration;
- output directory;
- timeout, retry and failure fallback;
- cloud shutdown behavior when applicable.
A previous render does not authorize a new render. A confirmed purpose, input, scope, output and fallback does. Preserve that authorization through retries of the same job.
Route execution
- Local: use
comfyui-local-runnerwith an already running ComfyUI server and an API-format workflow JSON. - Cloud: use
autodl-app-instanceto resolve and start the exact instance, thenminimax-h3-cloudto inspect and run an installed workflow by exact ID.
Never infer a cloud workflow from an old display name. Never resubmit merely because polling failed; recover the existing prompt or job ID first.
Compare cost
Use cost = hourly_rate * runtime_minutes / 60. Compare cost per successful
clip, not only hourly price. Include boot, upload, queue and failure overhead
when measured. Read references/runtime-selection.md
for the decision rule and references/cost-benchmark.md
for an editable example.