rzgar

/

minimax_h3_fl2v_lightx2v_4step_int8-convrot_comfy

like 5

you're welcome @xueqing12 here it is https://huggingface.co/rzgar/minimax_h3_fl2v_lightx2v_4step_int8-convrot_comfy/tree/main/workflow

you need to download and install the custom int8-convrot lora loader node first, if you haven’t already

yes, I use custom int8-convrot lora loader node.
https://huggingface.co/rzgar/minimax_h3_fl2v_lightx2v_4step_int8-convrot_comfy/tree/main/minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_norefiner_int8convrot
but there are 3 loras. how to use them to get best result,
what differenet between these three loras

By default, we use the converted ones. As I've mentioned and provided demo in the model card, those LoRAs in folders are for specific scenarios and experiments. In some videos, the sound is just glitchy or broken,
in my testing, the 'norefiner' variant solved some audio issues where the 4-step LoRA couldn’t handle it.

The attention (attn) and MLP (has most influence on sound) are separated internals so you can fine-tune the LoRA’s influence more precisely.
In multilingual scenarios, lowering the attn LoRA can fix some problems with non-english words or dialogues.
The best way is experimenting with them. For example, try attn LoRA at 0.33 and MLP LoRA at 0.86, generate a video, then fix the seed or use only one of them to see what happens.

thx your share and answer. That help me a lot! have a nice day.

you're welcome, you too!