Skip to content

The Cost of a Local AI Video Is Per Shot, Not Per Model

A ComfyUI node pack published measured MiniMax H3 shot times on one RTX 4090: 0.6 minutes for 2 seconds at 608×768, 9.3 minutes for 10 seconds at 928×1152, and an out-of-memory at the top size. The table is more useful than any model ranking, because it tells you which number to cut.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A stopwatch and a short strip of film frames on a dark metal bench beside a small graphics card, lit by one hard side light
One card, one stopwatch, and a shot list that has to fit inside both.

Somebody published a render-time table this week instead of another ranking. The ComfyUI node pack Infinite Movie, which turns a chain of stills into one looping film, added a table of per-shot render times measured on a single RTX 4090 with MiniMax H3 at 8 steps. It is the most useful thing I have read about local video this month, and it is useful for a reason that has nothing to do with MiniMax.

Here is the table, minutes per shot, 4:5 format, 24 GB card:

  • 608×768: 0.6 at 2 s, 1.1 at 4 s, 1.4 at 6 s, 2.1 at 8 s, 2.8 at 10 s
  • 768×960: 0.8, 1.7, 2.4, 3.6, 5.0
  • 928×1152: 1.3, 2.7, 3.9, 6.0, 9.3
  • 1088×1344: 1.9, 4.2, 6.2, 9.5, and out of memory at 10 s

A film adds about 40 seconds once, for loading, frame interpolation and joining.

The curve is the product, not the model

Read the table across, then down. At 608×768, going from 2 to 10 seconds multiplies render time by 4.7. At 768×960 the same jump is 6.25 times. At 928×1152 it is 7.2. The longer the shot, the more each added second costs, and the penalty grows with resolution, because the two multiply. A ten-second shot at 928×1152 is seven times the work of a two-second one, and at 1088×1344 a ten-second shot does not exist at all.

The README puts the shape down to attention being quadratic and notes that time grows faster than the pixel count. Whatever the mechanism, both resolution and length drive a shot's cost and neither scales linearly, so both numbers have to be chosen together. Deciding on 4K because the delivery spec says 4K, then on ten seconds because the beat feels long, is how a one-night render becomes a three-night one.

The top row is the part to take seriously. 1088×1344 renders at 8 seconds in 9.5 minutes and then fails completely at 10. Not slow. Out of memory. A shot list that assumed the model's advertised 10-second limit will not degrade gracefully at that size, it will stop.

Two engines, two different curves

The same card, warm, with other engines. Each was measured at its own size, so read the sizes before comparing:

  • LTX 2.5 distilled at 8 sigmas, at 768×960: about 1.1 minutes at 2 s, 1.3 at 5 s, 2.0 at 8 s. Its cost barely moves with length. MiniMax beats it at 2 seconds and loses from roughly 4 seconds on.
  • Wan 2.2 with the lightx2v LoRA at 4 plus 4 steps, at 608×768: 0.9 minutes at 4 s and 3.7 at 10 s, which the README calls close to MiniMax at the small size. But it loads two 14B experts, which adds about a minute before the first shot.

That last detail is the one people forget when they compare per-shot numbers. A fixed load cost does not matter on a forty-shot film and it wrecks a short one. If your film is four shots, a minute of model loading is a quarter of the render.

The node pack's own design also assumes this. Shot-level choice happens before assembly: you render, look at the shots in a contact grid, keep or redo each one, and only then does it build the film with frame interpolation and evened-out motion. Marked shots re-render alone. That is a review gate in the middle of the render, and it is the difference between a render farm and an editing process, because the expensive unit becomes the shot you rejected rather than the whole timeline.

Estimate from your own machine, not from mine

The detail I would steal from the project is the progress bar. It estimates remaining time from the durations of shots already rendered on that machine, stored in a local timing file, and then refines from the shots of the film in progress.

Estimates that come from somebody else's benchmark are wrong in a specific, expensive way: they are usually optimistic, and the error shows up as a shot list that does not fit in an evening. One memory measurement, one stored history, one estimate that updates from your own runs. Any local pipeline doing more than a few shots should be writing that file.

Until it exists, spend three shots on it. Render two seconds, six seconds and ten seconds at the size you actually intend to deliver, time them, and multiply. That calibration costs fifteen minutes at most sizes, and it tells you whether the film is a two-hour job or a two-night job, which is the only question that matters before you commit to forty shots. If the longest test runs out of memory instead of finishing, that is the answer too: the card has told you the shot has to get shorter or smaller before you can quote anything at all.

The card keeps getting smaller jobs

There is a second trend in the same week's Japanese roundup that will invalidate these exact numbers without invalidating the shape. A compressed MiniMax H3 variant, HyperFlow H3 Flash, is reported at 14B instead of 31B, generating video with synchronized audio in 4 to 8 steps, with a claimed 8.7 seconds for a 5.2-second clip on four B200s. The post names the 3090, 4090 and 5090 as supported, and a reply mentions a 4070 Ti Super working. Those are claims from a social post, not a benchmark of mine, and the workflow around them is not described.

Take the direction rather than the numbers. Each generation of open weights moves the same shot list onto a smaller card, so any table you build has a shelf life of months. The relationships inside it do not change: resolution and length both cost, the penalty for length compounds toward the top of the table, the fixed load cost is paid once per film, and the review gate is where the money is saved.

The other item in that roundup is a headline claim I can only take at face value, since the details sit behind a membership wall: a new MiniMax image model making 4K stills in about ten seconds, usable in ComfyUI from release day. Same story one step up the pipeline. Stills are getting cheap enough that the stills stop being the constraint in a shot. Rendering motion is.

What this changes for planning

Nothing about which model is best. Everything about how you describe a job.

A client asks for a fifteen-second product video and the answer is not fifteen seconds. It is four shots of three to four seconds, at a size the card can finish, with a variant per shot so there is something to choose between, and one measured number that turns the whole thing into a duration you can quote. The model name goes in the credit. The table goes in the quote.

If you have already fought this on a different card, the interesting comparison is not the absolute times, it is whether your curve bends at the same place. The fastest way to find out is to render one shot at your largest intended size, at your longest intended length, and see whether it finishes. That single test rules out the half of your shot list that would have crashed at the end of the night.

For the earlier argument about attention fallbacks and why the fast path sometimes never runs, see the profiling pass on LTX 2.5. And if your card is smaller than 24 GB, the MiniMax H3 quantization notes are the place to start before buying anything.

Sources

Share this guide
X LinkedIn