All right, you cyclopean abnormalities!

This is part two, [insert generous quantity of tentacles here]. Over the past few days I have trained three LoRAs for underwater work:

Two T2I style LoRAs for Wan 2.2 (T2V), trained on stills, both already on Patreon:

• Caustics, stills with strong caustic light.

• Underwater, nude female divers in a pool and in open water.

One T2V motion LoRA for Wan 2.2 (T2V), trained on 34 short clips:

• Submerge, female nude, semi-nude and swimsuit dives, pool and sea. Various(!) angles.

All of it on an H200: 140 GB of VRAM, 2015.4 GB of system RAM. High end! Rented by the hour on RunPod, five hours for the low pass and another five for the high. Yes. Yes.

With motion LoRAs you cannot cleanly separate the movement, the HIGH, from the detail, the LOW, so here are both at once. Thank me later … or support me on Patreon, where you get all the files you need and the workflows to build these clips properly. I also used some of my own custom nodes from this repo:

https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack

https://www.patreon.com/c/polyhedron_ai

So, why bother? Because underwater is where a base model quietly falls apart, and it does so in very specific places.

Caustics: that moving net of light on skin, sand and pool tiles. A base model either forgets it entirely or paints it on as a flat texture that sits perfectly still while the body moves underneath. It is supposed to drift on its own.

Eyes: underwater an eye loses focus, and across frames the pupils wander off or jump. Part of that is not really the model's fault. At training resolution, in a bust shot, an eye is barely a dozen pixels wide, so there is not a lot there to get right. But you can teach a model what an eye underwater is meant to look like, and that gets you most of the way.

Hands: the classic weak spot on dry land already. Put them underwater, splayed, half-blurred, crossing in front of the body, and you unlock the full horror-show catalogue. Same for hair and fabric, which float instead of falling. Gravity is the one thing a base model is really quite sure about.

Nipples and see-through fabric belong on that list too, but those are already handled by some older LoRAs of mine on Civitai.

None of this gets fixed by prompting harder. That is what the LoRAs are for.

Anyway. Highly experimental! Still tricky. It is not so easy to get everything right on so many levels.

And, yes: I am still on Wan 2.2. Not stubbornness, really: you cannot get to know the architecture of a model as complex as Wan 2.2 while forever panting after one new model on the way to an even newer one. Most people drive it through I2V, handing it a finished frame to start from. I work T2V, out of pure noise — where the model has to invent the motion instead of inheriting it. That is the harder half, and by some distance the more interesting one.

Case in point: MiniMax H3 — and as far as I understand it, there is a licence agreement saying you are not officially allowed to run it locally in the EU anyway. Yes, that is where we are now. As if we cared. ;D Check it for yourself anyway.

Sources:

Licence: https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/LICENSE

Repo: https://huggingface.co/MiniMaxAI/MiniMax-H3

MiniMax on the territory clause: https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/12