Kijai

/

MiniMax-H3_comfy

like 182

When using the Lora the audio gets massively distorted. I've seen some other posts about this and lightning lora's. Any tips or advise?

Unfortunately, it's happening to me too, on a 5090.

When did you update your ComfyUI last? They had a fix for it yesterday I think. At least, I updated this morning (it broke MiniMaxH3 Cache node), the LoRA works.

When did you update ComfyUI last? They had a fix for it this morning I think.

i try ... one moment.

And i change to sample: er_sde

When did you update your ComfyUI last? They had a fix for it yesterday I think. At least, I updated this morning (it broke MiniMaxH3 Cache node), the LoRA works.

Just updated comfy to the master/nightly branch :) And did a new generation! Fixed it. Thank you :)

When did you update your ComfyUI last? They had a fix for it yesterday I think. At least, I updated this morning (it broke MiniMaxH3 Cache node), the LoRA works.

ComfyUI update works. Sounds a bit metallic, but google gemini means, with some more steps it should sound better.

1376 x 768 | 8 seconds | 180,7s generation time ... a game changer LoRA .. thx @Kijai !!!

As with most things diffusion... its like rolling 100 dice and only numbers between 463-487 is "good" 😀

Anyway.. Settings:
Model: fl2va_pruned_Int8_convrot
Lora: lightx2v_turbo_4step (STR=0.6)
Sampler: euler/beta
Steps: 10
Extra node: Spectrum Apply MiniMax H3
Total time: 12:39
Hardware: 4070Ti Super 16GB - 32GB system ram (Yes.. heavy swapping)

I tried variations of strength at 4-6 steps, but i was not too happy about the result. Its not that the audio is "bad", but it has a much harder time hitting the mark with characters you have SOME idea of how should sound. Probably nitpickers out there that would comment on Darth Vader's voice.. but hey 😣

Atleast 55-60% faster than regular 20 steps for me.

Sampler: euler/beta

Wondering why people are using samplers different than the default. Did ComfyUI make the wrong choice with res_multistep + simple in their template?

I have similar setup plus MiniMax H3 Mem Eff Sage Attention Patch and ModelSamplingMiniMaxH3 (a.k.a. sigmashift, don't know why they have different names).
Yep, noticed that the audio is a bit flaky with 4 steps; it reminds me of old stereo cassettes that got worn out and the sound volume started jittering between channels. Definitely benefits from more steps, especially for the reference model with voice cloning.

As I have a fast SSD, I'm using --fast-disk parameter. It makes the first load faster and I haven't yet noticed it reading the model repeatedly, and system RAM stays low then (as models are not kept there, only operational data to swap). But I have 3090 with 24GB VRAM.

Not sure about why people use different samplers than the default... other than that "Spectrum Apply MiniMax H3" node seems to fudge around with tensors or somecrap.
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3#supported-samplers

What works for one prompt may yield a better result with a different scheduler for another prompt.. (Again, pointing back to my remark about rolling 10 dice....)

I will actually try that --fast-disk parameter, as i have not tried it. I have models on 1 NVME and the windows swapfile on another NVME, in an attempt to atleast not read model + write swap on the same disk simultaneously.

I noticed that using the MiniMax-H3 Turbo Sampler (4-step) node (https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo) improves the sound on the lightx2v lora but reduces the image quality. So it would be great if @Kijai could come up with something with this code so that we could get higher quality sound, but so that it would not spoil the image that lightx2v lora gives.

I noticed that using the MiniMax-H3 Turbo Sampler (4-step) node (https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo) improves the sound on the lightx2v lora but reduces the image quality. So it would be great if @Kijai could come up with something with this code so that we could get higher quality sound, but so that it would not spoil the image that lightx2v lora gives.

It's all been already fixed yesterday in core code, available in the nighly code and very soon in next stable version release.

I prefer to avoid installing custom nodes at all cost, so this lora is GOLD

With the most recent Comfyui update, I tried er_sde + sgm_uniform and lora weight 0.6 at 6 steps, the visual and audio quality are the same as offical 20 step workflow with much better motion.

As with most things diffusion... its like rolling 100 dice and only numbers between 463-487 is "good" 😀

Anyway.. Settings:
Model: fl2va_pruned_Int8_convrot
Lora: lightx2v_turbo_4step (STR=0.6)
Sampler: euler/beta
Steps: 10
Extra node: Spectrum Apply MiniMax H3
Total time: 12:39
Hardware: 4070Ti Super 16GB - 32GB system ram (Yes.. heavy swapping)

I tried variations of strength at 4-6 steps, but i was not too happy about the result. Its not that the audio is "bad", but it has a much harder time hitting the mark with characters you have SOME idea of how should sound. Probably nitpickers out there that would comment on Darth Vader's voice.. but hey 😣

Atleast 55-60% faster than regular 20 steps for me.

I have a almost identical setup except I have 96gb ddr5 ram and a 12 gb 4070. Could I please ask what if anything you have done to optimize your comfyui for low vram? Any launch flags? Other things?