alibaba-pai

/

MiniMax-H3-Acc-LoRAs

like 182

Follow

Alibaba-PAI 5.12k

what is needed to make these comfy versions?
what is being changed? how?
shouldn't there be some official documentation for this?

I thought they can be used directly with comfy . Can it not?

Didn't work with script, sorry.

Worked for me so I pasted the script here.

It does not work, guys. It uses https://arxiv.org/abs/2607.26004 which comfy does not support
Someone needs to code a custom node for it

You can prune these modules to run natively, but then it'll be a different LoRA. No one can tell if it works just the same

It does not work, guys. It uses https://arxiv.org/abs/2607.26004 which comfy does not support
Someone needs to code a custom node for it

You can convert or prune these modules to run natively, but then it'll be a different LoRA. No one can tell if it works just the same

Yes guys, it doesn't work like that, I closed my script thing too. Sorry.

https://huggingface.co/aptech0081/MiniMax-H3-Acc-LoRAs-ComfyUI

4070 32G：
4-step acceleration lora 5-second video takes about 11 minutes; ACC lora 5-second video takes about 20 minutes

Fixed in this Comfy PR:

https://github.com/Comfy-Org/ComfyUI/pull/15908
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

i tried that PR and got smeared hands, legs, bodies in fast motion scenes

i tried that PR and got smeared hands, legs, bodies in fast motion scenes

what about this one? did you try it ? https://huggingface.co/aptech0081/MiniMax-H3-Acc-LoRAs-ComfyUI

I tried the fl2va lora and used it for 4/8step T2VA generation. I find the 8 step results look good, although I met some blur hands/ layouts in high-dynamic secnes.

But when I trid 4 step inference with the same lora weight, the videos are not clear and seem not good.

I wonder why would this issue happened? I thought the lora should support 4step generation as it said.

i tried that PR and got smeared hands, legs, bodies in fast motion scenes

what about this one? did you try it ? https://huggingface.co/aptech0081/MiniMax-H3-Acc-LoRAs-ComfyUI

I did, I haven't tried it with recent comfyUI update
, but I doubt it'd change anything. Both LoRA present ghosting in fast action.

deno - has a custom node to load it. easy find in comfy manager