Add a reference and H3 runs in that mode; leave it empty and it falls back to text-to-video automatically. Guards flag oversized images (> 2048px / 8 MB) and clips over 15 s — placeholder limits, set from H3's real input constraints once the worker is serving.
The video engine this recipe runs on. MiniMax H3 today; more models (e.g. fal-hosted engines) drop in here later. The H3 task — t2va / ref2va / fl2va — is picked automatically from your reference above.
Uses a small/cheap model + your MD style guide to produce an H3-optimized prompt. Both raw and rewritten text are stored in the recipe.
Every result is a complete recipe — prompt, references, seed, settings and the worker that ran it. Re-run exact replays it identically; Duplicate & tweak loads it back into the editor with locked areas carried over.
Markdown the small model follows when rewriting your raw prompt into H3's structured format. Edited here, stored per recipe.
A preset is a saved partial recipe — your prompt areas (locked state included) and current settings, minus the seed and references.
Leave blank to run against the in-browser mock board (simulated workers, saved locally). Set your backend base URL to post real jobs — the app calls POST /generate, GET /jobs/:id, GET /gens against it.