Rkss/Minimax_h3_fl2v_lightx2v_turbo_4to8step_768p_v4_step600_dareties_native

silveroxides's MiniMax H3 Turbo LoRA (Experimental)

Experimental native-pruned version of SilverOxides' MiniMax H3 Turbo LoRA for ComfyUI.

Minimaxh3fl2vlightx2vturbo4to8stepv0.1-v1.0768pv4step600dareties_fro095

Load with ComfyUI's LoRA loader node.

Two versions. Pruned Adaln modules and a ported one.

This is not mathematically equivalent to the original LoRA

Original LoRA

Source: https://huggingface.co/silveroxides/MiniMax-H3_tests/tree/main/experimental

Recommended strength

• —FL2VA: 0.80 – 1.1
• —ref2VA: 1.0 – 1.1

Low strength may cause artifacts, softness, or voice instability. Use when you see oversharpen/detailed

High strength may become overcooked. Use when you see blurry ghosting.

Results may vary between workflows.

Recommended settings

• —Steps: 4-8 . It does work with 4 steps, but higher is better
• —Sampler: Euler
• —Scheduler: You can experiment with Beta57
• —If you want to use it with ref2va. Use the hybrid fl2va-ref2va model; recommend b20-49. https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models

Notes

This is an experimental build.