Experimental Turbo-Lora Blends/Tunes

A small comparison over some 4step loras: https://civitai.red/articles/36163/minimax-h3-turbo-lora-comparison

PDMD-DMAD-4step R256

PDMD is also new 4step distillation with good visual quality and slightly better audio over DMAD, but lacks some motion consistency and visual fidelity compared to DMAD, so this is my try to compensate for that.

⚡4 steps (possibly more)

⚡Euler + simple / Euler + beta

⚡Shift 12/3

DMAD-Hyperflow-4step R256

DMAD (used extract by Silver) is a new 4step distillation with stable visual quality and reasonable audio.

I found that the addition of hyperflow stabilize the motion depending on the scene and can normalize the gritty or very loud audio from DMAD.

⚡4 steps (possibly more)

⚡Euler + simple / Euler + beta

⚡Shift 12/3

Turbo-Multistep-Version 2

This Lora is mainly crafted from the hyperflow 1.0 , with fro 0.995 SVD and balanced consensus with factorized EMA600 turbo with tested and recalculated rank to minimize the the relative error median and stabilize visual and audio fidelity.

Hyperflow level's up the motion capabilities of H3 turbo by a lot and shines with ref2va.

EMA600 seem to compensate the visual and audio loss of hyperflow.

⚡4-8 steps

⚡Shift 6/3

Works really well on my tests with referenced audio, all modes (fl2va, i2va, ref2va) and all models including hybrid models.

Turbo-Multistep-Version 1

This Lora is a turbo (speed-up) lora for MiniMax H3 with multistep functionality, enhanced sound and awesome adherence and visual fidelity.

This LoRA is crafted/derived from the awesome LoRAs made by https://huggingface.co/silveroxides – support them if you can! They are awesome! 🤘

I used the experimental LoRA here: https://huggingface.co/silveroxides/MiniMax-H3_tests/tree/main/experimental.

Credits go to their work!

Thank you Silveroxides for your amazing work!

What I did and why this is different to this lora's:

I needed the lora highly compatible to my checkpoint and hybrid model without errors and adopting to the hybrid model.

I merged the lora into the MMH3 non-prunded base and made an SVD/Fro extraction 0.990 while recalculating the look-up table for the pruned model keys to exactly match the merge and pruned version. Capping this to rank 512. - This created a fork of the lora with a recalculated pruned version.

The lora comes in multiple ranks for different VRAM use-cases. I re-ranked the r512 base I created with SVD/Fro to:

r48 ~ 0.8 GB

r98 ~ 1.58 GB

r144 ~ 2.38 GB

r512 ~ 4.78 GB

Effectively always doubling the size and raising the overall VRAM consumption.

On my tests:

Rank 512 is highest quality. Every rank lower will work and the differences are only slightly noticeable, but on the lower ranks fine details or micro-motions may be not as good as on the higher ones.

Feel free to experiment with settings and use-cases. If you find anything noteworthy: I would like to hear! 👍

⚡4-8 steps (possibly more)

⚡Euler + simple / LCM + simple

YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.

⚖ Disclaimer & User Responsibility

All outputs, and all use or sharing of this file, are your sole responsibility. Outputs are machine-generated. I do not support or take responsibility for illegal, harmful, or harassing uses - and I will report content I become aware of that violates any license or platform terms of service.

THIS FILE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
I am not liable for any damages arising from its use. I am not affiliated with MiniMax.