Alright - so here's the software I've had GPT Astra build to make my life easier while working with MiniMax H3.

https://mega.nz/file/VptiUCYJ#yS1ouQryH4aLxVVXKl14N4cEZirvIyNad6p62eCAPzk
Minimum requirements: 16GB VRAM for the loaded models (there's still a lot of RAM swapping, so at least 32GB of RAM as well).
Runs on windows.

You start it via the Start-MotionBench.cmd file. Scan the archive with defender before running anything - as a good practice; I'm not giving anyone malware, but you never know on the internet.

It will run a script the first time it starts, installing a mini python runtime for the other scripts needed. If it doesn't find ComfyUI or it's other dependencies, you'll have the options to install and download all requirements - or point each dependency path to a file on your system.
You can view settings at http://127.0.0.1:8788/settings.html after the server starts.

The recommended model it runs is a version of encoder that allows for inputs of a starting frame + multiple reference images. You can then reference each image by mentioning 'Picture 1', 'Picture 2', etc. Picture 1 is always the starting frame.

The main things I like about it are that you can use the last frame of a finished generation as starting frame for the next - with the press of a button. A clip has gone off the far end from a specific point? Go to assemble with just one clip selected on the timeline and save a trimmed version up to that point, so it's not 100% wasted GPU time. Use last frame before things went off the rails and continue.

Got all the clips and story? Hop on the assembler, place them in order and it should be done.

There are some improvements I still need to make for it to be more user friendly, but I think it's quite useful with what's baked in already. Going to add some background removal via another small neural net so you can select reference images so background removal runs more easily.

Huh... my first contribution to open source if you can call it that, and it's a tool for generating smut.