QrusherZA/H3_LoRA_Resize_Experimental

Using the LoRAs in ComfyUI

1. Download the variant you want to try and place it in ComfyUI/models/loras/.
2. Load your MiniMax H3 checkpoint as you normally would. Connect its model output to a Load LoRA Model Only node, or use the LoRA loader that comes with your H3 workflow.
3. Select the adapter and run a short test clip. To tune strength, compare a baseline at 0.0 against a few positive strengths with the same seed, prompt, and sampling settings.
4. Check the ComfyUI console for unmatched LoRA keys. The checkpoint and loader must recognize the adapter's tensor names for the weights to apply.

MLP-only variant

The MLP-only variant applies the slider to the selected feed-forward layers. The resize recipe uses energy-based rank selection with a target of 99.5% of the original singular-value energy, a rank floor of 2