# MAI video tools (alpha) Three nodes and one widget that remove drudgery around renders without taking any decision away from you. None of them runs a model. ## MAI Video Out (file + sidecars + draft preview) Save Video with the record beside it. Frames (+ audio) go through ComfyUI's own encoder (h264 / av1, CRF), exactly like Create Video + Save Video; the extras are each off unless wired or toggled: - `hold_map` -> `_00001_.holdmap.json`: the clock that produced these frames. H3 Load Hold Map and the deck's hold-map staircase read it. - `write_meta` -> `_00001_.meta.json`: seeds, models, LoRAs, steps, samplers, inject schedule, encode time, your `notes`, all read out of the graph that ran. The compare viewer's metadata strip and the scorer's labels come from here instead of from file names. - `draft_preview` + `latent`: decode the sampled latent through the tiny VAE in `models/vae_approx` (`taeh3` for H3; any stem) into `_draft.mp4`. The tiny VAE is loaded for that call and released after; with the toggle off nothing is loaded and nothing runs. Cost measured 2026-08-21: 8 ms per latent frame at 864x480, 21 ms at 1344x768, 42 ms at 1080p. ## MAI Load Video (path) / MAI Select Every Nth The two things the graphs still took from VideoHelperSuite: a video by absolute or input-relative path (as ComfyUI's VIDEO type, frames and fps alongside), and every n-th frame from an offset. For a smeared clip use H3 Exact Recover instead; it is exact. ## MAI Video Compare (2-6, synchronized) The player has grown since this page was written: a waveform timeline with loop brackets, a frame stepper, and a frame-exact export. The README's MAI Video Compare section is the current reference; this page keeps the original wiring notes. A media UI, not a processing stage. The node writes each wired VIDEO once as a small h264 preview into the temp directory (CPU, no VAE, no tensors kept) and the widget plays them in the browser with the flipbook's player: one set of live `