TL;DR: This LoRA is made for goth girls and cleanly overrides everything, even with accelerators active, including the dreaded plastic skin. A weight of 1.0 is fine and gives good results without any other LoRAs.

------

So, creatures of the night,

back to my roots, my beloved goth girls, this time in MiniMax H3. <3 For this I trained a new and fairly strong LoRA that also gets a firm grip on the accelerators, especially when it comes to that plastic skin. Set the LoRA to 1.0 and you get a complete override.

Speaking of accelerators: far more important than the looks are the updates to my Polyhedron Suite, a custom node pack for ComfyUI. You can grab the current version here:

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

The pack has been available through the ComfyUI Registry for a long time, but a few newer versions are still waiting in the review queue. If the ComfyUI Manager offers you an older one, skip it and take the current version straight from GitHub.

The Suite now supports "minimax_h3_hyperflow_8step" with its own sigma node, as well as the "MiniMax-H3-Ref2VA-Acc-8Step" and "MiniMax-H3-FL2VA-Acc-8Step" models, with warnings if something is set up wrong. That brought my render time down to 4 minutes for 90 frames (3.75 seconds), not counting the ESRGAN upscale afterwards. That one eats another 8 minutes, but it now works properly as a real refiner.

(My tech gear: RTX 5060 Ti with 16 GB VRAM and 32 GB RAM)

My main model right now is "Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8", which is very good, together with several custom LoRAs next to this one to push everything a little further.

You'll find the workflows, more examples with prompts and all the LoRAs I used on my Patreon page. So if you like what I do, a little support would be great:

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

Other than that: have fun with this LoRA. The world needs more goth girls! :D

Take care,

Frank