Version 1
This Lora is a turbo (speed-up) lora for MiniMax H3 with multistep functionality, enhanced sound and awesome adherence and visual fidelity.
Credit
This LoRA is crafted/derived from the awesome LoRAs made by https://huggingface.co/silveroxides – support them if you can! They are awesome! 🤘
I used the experimental LoRA here: https://huggingface.co/silveroxides/MiniMax-H3_tests/tree/main/experimental.
Credits go to their work!
Thank you Silveroxides for your amazing work!
What I did and why this is different to this lora's:
I needed the lora highly compatible to my checkpoint and hybrid model without errors and adopting to the hybrid model.
I merged the lora into the MMH3 non-prunded base and made an SVD/Fro extraction 0.990 while recalculating the look-up table for the pruned model keys to exactly match the merge and pruned version. Capping this to rank 512. - This created a fork of the lora with a recalculated pruned version.
The lora comes in multiple ranks for different VRAM use-cases. I re-ranked the r512 base I created with SVD/Fro to:
r48 ~ 0.8 GB
r98 ~ 1.58 GB
r144 ~ 2.38 GB
r512 ~ 4.78 GB
Effectively always doubling the size and raising the overall VRAM consumption.
On my tests:
Rank 512 is highest quality. Every rank lower will work and the differences are only slightly noticeable, but on the lower ranks fine details or micro-motions may be not as good as on the higher ones.
Feel free to experiment with settings and use-cases. If you find anything noteworthy: I would like to hear! 👍
Facts:
⚡Multistep capability
⚡4-8 steps (possibly more)
⚡Stable sound
⚡Stable visuals
⚡Works with reference audio, video and image files
⚡Works with fl2va, ref2va and hybrid
⚡No noticeable style or lighting drift
⚡No visual morphing
⚡Pruned compatible
Settings I tested:
⚡4-8 steps (8 recommended)
⚡Euler/simple
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
⚖ Disclaimer & User Responsibility
All outputs, and all use or sharing of this file, are your sole responsibility. Outputs are machine-generated. I do not support or take responsibility for illegal, harmful, or harassing uses - and I will report content I become aware of that violates any license or platform terms of service.
THIS FILE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
I am not liable for any damages arising from its use. I am not affiliated with MiniMax.
Description
V1
hybrid multistep
4-8 steps
Base: multistep_v4_step600_delta_cwb_7_contrib_bf16
pruned
re-calculated
FAQ
Comments (44)
Is this LoRA used with the v1 model? In fact, I found that the base v1 model can also produce fairly good results in 8 steps without the LoRA.
This Turbo lora really made a huge difference in my videos. Wow ! :O Thanks for sharing it !!
seems understand prompt well, mixing this with light and ema600 results better
how so? what values do you have your strengths at?
I would be interested in that mix, if it can enhance the recipe :)
@Darksidewalker I'm playing around with larry's at .3 to .48 and light ref2va 768 V1 at around .2 to .5 and yours with .5 to .8
lovin this lora though man. You really made it possible for sound and visual to be amazing! Thank you - where's the tip jar?
@A_fuzzy_larkin Thank you! The original lora is from silveroxides!
I made some changes to it (like described), so maybe they could use the tip more :)
If you really want to tip me, my ko-fi link is right in the description (image) and the profile.
@Darksidewalker perfect! I'll tip both of you for the amazing work!
After doing a few runs with the r144 version , I believe is the best results I have had so far with any turbo lora , amazing job as always!
Do you see any major differences between your v1+this lora vs your v1 4/8 step built in?
I did not try it with v1. So I cannot tell the diff there, sorry.
Do you recommend v2 full+this lora or just v2 turbo directly?
Extra lora needs extra VRAM. If you need/want turbo baked is better for vram saving, if you want more control, extra lora.
@Darksidewalker got it, thanks a lot.
great work bro.Finally..
what min config do you recommend with your lora ? I struggle to run it with a 5070 Ti
There are multiple uploads in the dl window, should be something for every vram. filesize ~ vram usage, round about.
no difference ....
Love your work. Thank you for this.
Can you explain what's going on with the file variants?
The visible part of the file name says bf16 for all of them.
The first three are tagged as fp32 and the last one agrees with the file name as bf16.
They're all different sizes and the one tagged bf16 is largest. What is the actual difference between them?
You can hover your mouse over the name to see the full name. The tagging is not made by me, it was civitai automatically. And it is wrong.I'll write it into the description what file will be what rank.
@Darksidewalker Understood. Thank you!
DaSiWa, holy fuck... You're incredible.
This lora could render clearer fingers and toes if your hands and feet were in the frame, but will result in extra fingers or toes if hands and feet initially out of frame re-enter the frame.
Oh really? This was not happening with my tests. Something to look at for a next version.
Somehow, it works really well with v1 dasiwa, but I just couldn't get it to work with v2. When paired with v2, the outputs were all half-rendered, no matter how many steps I added.
I only used and made it for v2, I never got "half-rendered", can you explain what you mean and what you did?
The output videos were all blurry and ghosting, looked like they needed more steps to be fully rendered. I forgot to mention, I am doing 2 passes workflow, from 0.4 megapixels to 0.7 megapixels. I tried adding more steps in the 1st pass, in the 2nd pass, in both passes, and even bumped the lora strength to 2.0, but no luck. With the v2 turbo model, the same workflow worked, so I was very confused why v2 model + this lora, would not work.
@john2stai I can assure, with my workflow the lora works just fine with every model I tested. This must be an issue of your workflow somehow.
Ah Dasiwa, when i see your name i first click download and after that i look what i am actually am downloading. it is just that simple if you made it than i want it. thanks for all your work
I noticed the title includes REF2VA | FL2VA | Hybrid, but these five files only mention FL2VA. Are you planning to release the remaining two versions later, or does this 5.72 GB file include all three?
Sorry, it’s 4.78 GB.
It is like described
Does a fantastic job, I don't know if it makes sense but in combination of the minimax h3 turbo v4 step600 lora it gives even better results and better image quality.
Thank's, what is this combination, any details for testing?
@Darksidewalker I have both lora's active, one is the minimax_h3_turo_v4_step600_pruned on strength 0.75 (downloaded it from civit ai red) and the other is this lora the minimax_h3_fl2va_b16_turbo_multisptep_fro099 on 1.00 strength. I am using both these loras on 8 step video generation and the result is crisp video and audio.
@sarahbonestone that would be distillation on 1.75 strength. Seems a bit high? What do you think is working better with that high strength - is there a specific use case? Or overall?
@Darksidewalker I am not sure, I don't know what does high distillation mean, I accidentally stumbled on this one but provides me the best video and audio results. Overall gives me great results. Maybe I should reduce the strength of the 1.0 lora to lower, need to experiment it more.
@sarahbonestone I'd be interested in some AB testing, for sure.
@sarahbonestone turbo = distillation, less steps to an outcome.
But that much strength on a distillation process normally has downsides. But I did not test that high.
@Darksidewalker I am far for an expert so I use comfyui at relative beginner level, all I know that I these two loras turned on for all of my video generation and so far it provides the best results for me. As I mentioned I use currently minimax_h3_turo_v4_step600_pruned on strength 0.75 and the minimax_h3_fl2va_b16_turbo_multisptep_fro099 on 1.00 strength. I think that the second one can be reduced from 1.00 to lower and the results will still be great. I accidentally found out that these two loras create great fast results. I always keep 8 steps in my KSampler, tried 10 it lasts a bit longer but didn't noticed much improvements. Have not tried to go below 8 step, maybe it does great at going a bit lower I don't know.
@sarahbonestone You only use 1 speed lora at a time. If you dont have good results with 1 speed lora and you need 2 you doing something wrong somewhere else. But as you said your a beginner and still learning. Thats why im telling you to use only 1 speed lora and check your other settings. Also test different workflows. It needs alot of time and testing to become good at this stuff.
People might think you're tripping but I tried this myself, running the larry 600ema at 0.7 and the fro99 version of this one at 0.6 and it is making everything look crystal clear and sharp. Faces in the distant background still look weird like usual but everything looks high quality even when I have characters moving a lot.
I've seen other people combining turbo loras before, usually at 0.3-0.4 for each of them using up to three different ones for whatever results they're looking for. So it isn't unheard of.
I'm using Wangp which does a lot under the hood with the loras already to make them work with any model so maybe that's doing something too but I'm definitely getting better results using this lora with the larry lora.
@RenegadeZebra Exactly, I have no explanation and maybe it makes no sense, but since I discovered I keep both loras the h3_turo_v4_step600 and the fro99 turned on for all for my generated videos and it gives as you said crystal clear and sharp results.
Can I use it with original minimax_h3_fl2va_int8_convrot?