A skin realism LoRA for HyperFlow H3 Flash (ref2va and fl2va) and MiniMax H3 video: less plastic shine, more natural skin.

What's new in V2

V1 was a block-renumbered port of @alcaitiff's original rank-1, MLP-only slider (https://civarchive.com/models/3014078). Because V1 only touched MLP layers, smooth reference portraits in ref2va passed through it by way of self-attention and barely changed.

V2 is trained from scratch at rank 16 on both self-attention (qkv_proj, out_proj) and MLP (fc1, fc2) in H3's later blocks (100 layers):

• It changes smooth reference skin in ref2va. It takes off plastic highlights and brings back natural skin texture. We tested it on fair, olive, East Asian, South Asian and deep-melanin skin, and did not see added grain or ageing.

• It is trained to leave fine features alone. Eyes, eyelashes, irises, eyebrows, beards and hair stayed sharp in our tests, in motion and in 1080p finals.

Which file to download

• V2 - ref2va: for minimax_h3_ref2va_hyperflow_flash_14b with ComfyUI's stock LoraLoaderModelOnly. Strength 0.65 to 0.85, default 0.75.

• V2 - fl2va: for minimax_h3_fl2va_hyperflow_flash_14b with the stock loader. Strength 0.65 to 0.90, default 0.80. 97 of its 100 layers load there; the warning about the other 3 is expected.

• V2 - Base H3: the file for H3 Editor (comfyui-h3-editor, https://github.com/igorls/comfyui-h3-editor) on any checkpoint, at strength 1.2 with start_sigma 0.75, and for the stock loader on the unpruned 50-block MiniMax H3 at 0.65 to 0.85. With H3 Editor, use this file even on HyperFlow: its H3 LoRA node renumbers the blocks itself.

No trigger words.

The comparison clips are short, one to two seconds each, and show the same shot with and without the LoRA.

(The V1 tabs stay available. Credit to @alcaitiff for the original V1 slider.)