V0.2 has 2 variants because I couldn't choose between the two. Choose epoch 6 or epoch 7. There's a I2V comparison for V0.2 variants in the Frieren example, it is not a comparison of v0.2 and v0.1. Optimization nodes might heavily affect output.
Put something like this in your prompt:
twerks causing her ass to jiggle and shakeSettings I somewhat tested:
1 STR for MAXIMUM JIGGLE. (may have to turn it down when mixing loras? idk)
0.5 STR if you don't want Asians in T2V. Too high will ignore your race/skin color prompt but shouldn't be a problem with I2VA. Using optimization nodes affects outputs motion/speed I believe.
I have not tested REF2VA so feel free to try it out and share results! I did not really seed hunt for the examples so I could upload this model asap. I trained at 3000 steps and 10 epochs. Thanks to anonimoose for sharing the dataset.
Workflow embedded in some vids (be aware of LoRa name mismatch in the lora loader) OR here to stay up to date w/ workflow (recommended and may have to install custom nodes/update ComfyUI.)
MiniMax H3 slop breast bounce lora here, might work together?
https://civarchive.com/models/2876282/slop-bounce-minimax-h3-lora
MiniMax H3 Prompt guide for you/LLM.
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
If body is too big, try pairing with a breast/body type lora and slightly lower strengths.
If clothes are vanishing, you may have to prompt clothing.
Use responsibly. Please follow Civitai's Terms of Service and applicable laws.
Description
Put something like this in your prompt:
twerks causing her ass to jiggle and shakeSettings I somewhat tested:
1 STR for MAXIMUM JIGGLE. (may have to turn it down when mixing loras? idk)
0.5 STR if you don't want Asians in T2V. Too high will ignore your race/skin color prompt but shouldn't be a problem with I2VA
This is an early test LoRa (Epoch 3) from the initial training run. The run ended unexpectedly before reaching the planned final epochs, so this may not represent the LoRA's full potential. A new training run will be attempted for a future version. Quality starts getting good by the third epoch anyways. I noticed improvement in physics with T2VA/I2VA vs no LoRa. Have not tested REF2VA so feel free to try it out and share results! I did not really seed hunt for the examples so I could upload this model asap. Enjoy this early one for now. I trained at 3000 steps and 10 epochs. Thanks to anonimoose for sharing the dataset.
Workflow embedded in Frieren vid (be aware of LoRa name mismatch in the lora loader, so you have to add it like you would normally) OR here to stay up to date w/ workflow (may have to install custom nodes/update ComfyUI.)
MiniMax H3 slop breast bounce lora here, might work together?
https://civitai.red/models/2876282/slop-bounce-minimax-h3-lora
MiniMax H3 Prompt guide for you/LLM.
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md
FAQ
Comments (24)
This is what we've all been waiting for!
Thanks again for the buzz <3
@SugarPixelArt thx for the continued support! I might have a new version of this lora up soon after testing :p
@bhopping You're welcome! Thank you for all the work you've put into this.
This is one of Minimax's most important LoRAs!
It's certainly going to "shake things up".
EDIT: OP fixed the description. Leaving the comment for the discussion below, in case it helps others. We were discussing how the DaSiWa LoRA Loader works.
How does it work then? I am new to it and relayed info of my findings with it.
I find VIS to be useful personally. I fiddled with it a bit and it seems to be how wan HIGH/LOW loras work if you like customizing of the LoRa. Like for example, using T2V at STR 1 and VIS 1 will produce Asians even when prompting a different race. Lowering just the VIS to 0.5 solved this issue adhereing to my prompt better about the visual of the people. Just like how I keep my motion LOW wan loras at 0.5. Try it yourself you'll see what I mean.
@bhopping Try the same thing with only "Str" at 0.5 and "Vis" and "A" at 1.0. Should give you the same result.
@zkyt Ok I will try that ty
@bhopping please let me know if it works! I would like to know if this is actually right or if I misunderstood something with the DaSiWa nodes
@zkyt That's wild because I initially tested STR 1 and VIS 0 then vice versa to see a difference and there was but now there isn't? I must've mess something up during testing and gave me the impression of how WAN worked just by coincidence. It's also funny when LLM's just blatantly played along with the idea, really made me think I was onto something there lol thanks for the info I will fix desc
@bhopping Yeah I got the same problem with LLMs sometimes :D
Thanks for updating the description and good work on the Lora btw!
@zkyt I think I got it now. So basically the VIS is just a multiplier for the strength itself. Also brother who made the node should've name the knob MULTIPLIER. “VIS” makes it sound like a special visual feature control. Like me.
@bhopping Yeah I agree, the node is very confusing with the descriptions.
Fantastic work!!
epic cool
When you train, do you use t2v or i2v as the base? So far I've only been trying i2v, but wondering if maybe t2v is better? My training attempts have been awful so far XD
I trained on Civitai and there's only one base model option to choose from called MiniMax H3 base so idk about that one. You have the option to still train I2V or T2V? I thought its FL2VA and REF2VA for MiniMax H3's base models. How many steps are you doing? Did you test each epoch in ComfUI too?
@bhopping ah I'm training on my pc, I have a 4090. My biggest issue was trying to get breast micro physics. But I finally got it, works really well with your twerk lora. I actually had to train a lora for the text encoder. Thanks!
THE GOAT
Thx for the buzz : )
The return of the slop twerk; God smiles upon us this day!