The full LoRA also patches the model's token_refiner - the text/audio
conditioning refiner. The norefiner variant removes those 8 patches so the
conditioning path runs at base quality (some video clips had bad audio. This solution fixed it, but keep in mind it doesn’t happen always).
If audio/text conditioning feels weak with the full LoRA at 4 steps, try the norefiner variant.
First clip: with refiner | last clip: norefiner
Usage (ComfyUI)
Unzip the ComfyUI-LoraInt8Loader custom node into custom_nodes, restart
In your MiniMax H3 workflow, add LoRA Loader (Int8-ConvRot), select the
file, strength 1.0 (the training scale is already baked in)
