Fizgig 9.6 adds LoRA training for MiniMax H3 Image, fits 16 GB cards

shootthesound · reddit · 2026-11-11

LoRA training tool Fizgig 9.6 adds support for MiniMax H3 Image (text-to-image plus edits with up to 16 reference images, 6 steps, no CFG).

Key points:

Trains LoRAs, LoKRs, edit LoRAs from before/after pairs, and slider LoRAs; masked training, Multi Concept, LoRA Isolation and Studio all work
The author trained a training adapter that prevents distortion when training this model — on by default, never saved into the LoRA, and free on Hugging Face
Practical guidance: prefers quality datasets, presets start at 0.5 MP; 23 GB VRAM at int8, 13 GB at 4-bit; a 16 GB card can train and run Studio
ComfyUI gotcha: its default dynamic VRAM loading adds LoRAs in bf16, rounding away 15% on this model; --disable-dynamic-vram restores full strength — reported with a suggested fix
General improvements: automatic OOM recovery (up to three tries from the newest save), failed runs no longer block the queue, AMD installs now use stable ROCm PyTorch

More from Multimodal