Limits and performance#
Envelope#
| LTX | MiniMax H3 | |
|---|---|---|
| Frames per take | up to about 257 | 17n + 5, up to 362 (15.1 s) |
| Audio | up to about 10 s stays clean | full take, speech and ambience |
| Resolution sweet spot | preset native | 768 square |
| References | first frame | up to 9 pictures (5 tested: two faces, a location, a pose guide and a composition guide) plus voice references |
| VRAM | 16 GB | 16 GB with int8 packs |
| Host RAM | 32 GB | 32 GB minimum, 64 GB recommended |
H3 takes have their audio normalised to about −16 LUFS integrated loudness (true peak −1.5 dBTP), so consecutive takes play back at a similar level.
When the runtime releases its warm pipeline (an image job or the assistant needs the room), it gives the memory back: the idle runtime keeps about 5 GB of host RAM, against about 20 GB before 2026-10-23.
Render time#
The first take after a start loads the pack (a few minutes). After that, a 768-square H3 take of 124 frames renders in a few minutes on an RTX 5070 Ti class card; a 15-second (345-frame) 1344×768 take with one face reference takes about 17 minutes with a standard int8 pack. Fast int8 packs cut each denoising step by about a third: the same take finishes in about 11 minutes. Both figures use the default match reference canvas; native adds about three minutes. The stock presets (Ref2VA and the text-to-video H3 preset) get the fast kernel through Speed up; at 768×512 the gain is larger: a 124-frame take measured 74 seconds instead of 160. Without it a catalog install loads the bf16 model file and quantises it on every cold load, so the first take after a restart starts a few minutes late. Every extra reference picture adds to the time: with five pictures (two faces, a location, a pose guide and a composition guide) a 124-frame 768×512 take measured 87 s instead of 69–75 s, its steps about 30 % slower. A 345-frame 768×512 take with two faces and two voice references took 3.5 minutes at a 6.7 GB peak. A clip chain resets the GPU state before every continuation clip; since 2026-10-08 that reset keeps the loaded weights and takes a fraction of a second (a three-clip 768×512 chain: about 3.3 minutes instead of 4.5). Keeping the references costs about 4 % more per clip.
Every H3 take starts by encoding the prompt with the 32B text encoder, which is read from disk each time; since 2026-10-23 that takes about half as long (a short prompt: 9–12 s instead of 18–24 s). The runtime keeps the encoded prompts, so a take with the same prompt and the same reference pictures (another seed, the other arm of a comparison) skips that step; about 2 GB of the most recently used prompts are kept. The video is the same either way.
When it is slower than that#
Another process holding GPU memory is the usual cause. Voice assistants, local LLM servers and speech tools can squat several gigabytes; Maystro then spills into shared memory and every step slows down by a factor of three or four. Close them before rendering; Maystro's own assistant model needs no attention, it is unloaded before every render and reloaded when the queue is idle. See Troubleshooting.
Cancelling#
Cancel from the take card or the queue. The runtime finishes the denoising step that is running and stops, so a cancelled take ends within one step: a few seconds on a short preset, up to about a minute on a long 1344×768 take. Only the model load and the final decode have no stopping point; a cancel that arrives then is honoured right after that phase. The take is marked cancelled and no file is written. The model stays loaded, so the next take starts without a reload.
Out of memory#
If a take fails with an out-of-memory error: shorten the take, drop to the 768-square preset, remove the extra reference images, and make sure nothing else is on the GPU. A 15-second 1344×768 take with one face reference fits the 16 GB envelope.