A port of [MMH3] Realism Slider by @alcaitiff to HyperFlow H3 Flash, the distilled MiniMax H3 GGUF. Same weights, no retraining: only the block numbers changed. Original slider: https://civarchive.com/models/3014078

Why it needs a port

HyperFlow pruned whole transformer blocks from MiniMax H3, and its two models pruned different ones. ref2va keeps 49 blocks (H3 block 9 removed) and fl2va keeps 47 (blocks 6, 21 and 22 removed). Loaded as-is, an H3 LoRA lands on the wrong blocks after each gap. These files have their keys renumbered for each model.

Which file

fl2va version: for minimax_h3_fl2va_hyperflow_flash_14b, prompt-only video.

ref2va version: for minimax_h3_ref2va_hyperflow_flash_14b, video with reference pictures or first and last frames.

Load it with a model-only LoRA loader after the HyperFlow GGUF loader. On ref2va all 100 weights load. On fl2va 98 of 100 load: HyperFlow stores block 46's second MLP layer in compressed form, so ComfyUI reports blocks.46.mlp.fc2 as not loaded. That is expected.

What it does

Tested on HyperFlow at 6 steps, 768x1344, with the same seed and prompt per character. Glossy, over-sharpened skin turns matte and natural, while pores and wrinkles stay. The effect grows steadily from 0.5 to 1.5; 2.0 adds little and can make fabric texture noisy. On HyperFlow the slider also shifts framing and background a little at the same seed.

Recommended strength: 1.0 to 1.5.

All credit for the slider goes to @alcaitiff