FinalCut MiniMaxH3 (Beta Lora)
Before commenting what does this lora do, inform yourself on what a finetune extraction lora is. I can not list the 100's or more of SFW and NSFW concepts the finetune has been trained on.
All concepts where hand captioned or corrected describing motion, action and frame changes.
This lora will fully change the behavior of Minimax to match the finetune within 90%
The Lora was trained on both SFW and NSFW concepts but trends toward NSFW and has been marked as such
Note: this lora is marked as Beta, and is not like any other Minimax H3 lora that I am aware of currently so I can not fully predict the effects, and it is still under testing.
Description
FAQ
Comments (64)
Any inventory or list of these actions to get a better sense of what it should be used for?
I can not list the amount of SFW and NSFW concepts it was trained on as that list would be to long - Also it is currently in BETA and being tested both onsite and offsite but both those things take time
@Felldude understood or maybe give the overall high-level concepts we should be trying have it influence as there's only one test video which probably doesn't do your hard work justice on what we can use it to influence, you know?
@Grockster Initial testing saw an increase in NSFW prompt understanding when using Ref and a starting NSFW image, or as the cover video shows a SFW image prompted into NSFW - It may also increase diversity in the SFW category but seems to be biased toward NSFW
Agreed. Would like to know what actual things it was trained on to know what to prompt for
I can not list the amount of SFW and NSFW concepts it was trained on as that list would be to long
Realistic Selfie Videos, dancing, youtube, X.com, 100's of concepts........It'd be nice if you listed some of the 100's of concepts before I start a 500+mb download.
If your in an area where a 500MB download is an issue, you will need to wait for me to finish testing offsite and onsite and take this LORA out of beta
More examples or information would be appreciated.
A short video of a woman shaking her butt infront of a mirror doesn't tell us much about what to do with this lora.
vague lora is vague
most trust me bro lora ever.
First of all, thanks for sharing your LoRA! If possible, I’d be really happy if you could list some prompts I could use with it.I'd prefer that be an onsite demonstration but the site is hours out on generation
You talked and talked but didn't say a thing... What does this lora actually do?
While I appreciate the questions regarding what to use to trigger this lora, that is not really applicable as the lora will fully change any text to video, image to video or first to last frame to align to my Finalcut training - It tends to shift toward NSFW concepts, this is amplified if text to video is used.
Note: Please note this lora is marked as Beta, and is not like any other lora that I am aware of currently so I can not fully predict the effects.
@Felldude - are you not grasping that the majority of people just have no idea what the "finalcut" reference is?
If you could just explain it with a single sentence and/or provide a link everyone would be informed and the slew of silly comments would cease.
Like ... the text as it is now essentially insists that I go do research to discover what your proprietary terminology refers to... and you have had plenty of time to simply explain it.
I'm busy hermaflating my buffalo dismount valve right now so I literally do not have time to do research on some proprietary -thing-.
Do you think you are famous and everyone knows your history and past accomplishments so you don't have to explain what you're talking about?
The huffy attitude about it is silly.
It's not about how to prompt the model, it's about what is the model supposed to do...
I`ll go against the stream of commenters here and download it and try it, hell - i might even google what "Final Cut" is and do some research and actually test the thing, cause here`s the thing: its FREE, I apologize on behalf of the other commenters their entitled behavior.
That I can answer without finishing my testing - FinalCut is the name I give to any model that is trained with a known large working dataset I have - The addition of of synthetic video and other unknowns gives this the BETA notice
Its bad behavior to ask what it does? lmao
we all look forward to your posts in the gallery
@SomeRandomUser23 Absolutely not, but it -is- bad behavior if a person "asks" simply in a way to in a toxic stabbing way pre-judge something as bad whom a lot of commenters have on this page without actually trying it.
it`s disgusting behavior, and defending it is equally so
@mmonkerud534 My god you are so dramatic
how much did your google search help you out there bud?
Wtf is wrong with people.. i LOVE mystery boxes, cant wait to A/B some seeds to see what different strengths of this lora does, hell wil be even cooler when i edit the weights. Thank you sir for a few hours of fun :)
and wat u get inside the mystry box, enlighten us plz
A/B even those that show failure are always welcome especially if the seed prompt is repeatable
wow this lora is amazing, it fixed all my problems in life and i now own a home and have a wife with two kids. thank you for this lora
See I had the same thing happen in reverse
Not only does he not directly explain what this LoRA does, but he also hasn't added enough example videos.
sure he did, he explained it as it's a full conversion of final cut. everyone knows what final cut is, right? 🤣
what, that girl bouncing up and down in front of a mirror doesn't explain everything?
I am more amazed that folks who've never made a lora before are somehow complaining....
(I know a few of you have...) not talking to you.
If you haven't, then maybe cut Felldude some slack. He is one of the few creators genuinely trying to move the ball forward for all of us in the AI image generating community.
Making even unappreciated loras is a task that takes time, coin, sweat, and tears and Civitai doesn't make it easy.
hi, lora creator here. not describing what your lora is and how it works is stupid. defending that practice is stupid. i sure hope you've never complained about a movie or game, and if you have, you better have years and years of experience in film-making and game design...
the only credit i can give to the creator is they aren't charging 2000 buzz for their mystery lora. it could be an amazing lora, but i have literally zero information to base its quality or effectiveness on.
As I have stated numerous times this is not a lora in the single concept sense this is a finetune extraction this is not the same
@Felldude no one wants you to write a novel about every little thing this lora was trained on. but "this lora will fully change the behavior of minimax to match the finetune within 90%" is about as empty of a statement as Taco Bells tacos. change the behavior in what way? will motion look smoother? will it follow my prompt better? could we get a broad idea of the concepts represented? the vibe of your description feels like "this lora makes the model different, and that is better." the lack of before and after previews also lends to this vibe.
And even so, I don't know what this thing does. I'm not going to waste time trying to guess what it's for.Right, so I downloaded this 'FinalCut', gave it a fair go, and… nah. Bloody disappointing, mate. The only thing that needs cutting here is this lora itself, straight out of CivitAI's storage. It's just wasting everyone's time and space. Don't bother. Shit sucks mate. It's overcooked and the character drifts harder than a Toyota AE86 Trueno on Mount Akina.
Overcooked is a possibility as the rank and alpha was fairly high - the other half of people will still claim this lora does nothing though
Holy fuck, Felldude. This is exceptional. Took a little bit of playing around, but I've never had character consistency, motion, and just general quality like this before.
People, what Felldude means by "fully change" is simply you type whatever and see how things go.
From my own testing, and this isjust a few generations, but between 0.3 and 0.5 when using other LoRAs. 0.6 and 0.8 when by itself or maybe one or two extra LoRAs. Honestly it just depends on the model and loras you're using. BUt in general, the results have been very good.
Bravo, Felldude. And thank you, I know it's beta, if you continue to develop this then we're all in for a treat.
Thanks, it is still in beta largely do how many ways you can input data into H3, text, ref, and FL2VA
So you mean it follows prompts differently? or the quality is different? I'm trying to understand how do use it. Have you simply applied it to workflows you've already done to compare differences?
Just so everyone is aware, this is a troll post. The user seems to enjoy uploading LoRAs with vague, nonsensical descriptions, seemingly to prompt confused users into asking what they do—only to receive irrelevant, non sequitur responses in return. He did it before with this one: https://civitai.red/models/2119634/consistent-face-wan-video-14b-t2v
That is one heck a of career of trolling. That account contribution is kinda large.
I agree that it's pure troll but it's pretty low stakes and nobody's feelings are getting hurt. I mean it's weird but he's only burning his own gpu hours.
prompting "hair cut" didn't do much either.
Why do some people always give LoRAs baffling names and then make the descriptions as cryptic as a riddler, never clearly explaining their specific purpose or how to use them? You worked hard to train a LoRA precisely because you want more people to use it and appreciate it, right? You can train a LoRA, yet you can't come up with an intuitive, good name? Maybe this LoRA will later prove to be useful, but this approach is very likely to bury it.
was it like SecretSauce,a lora used to remove clothes without change the origin video.
I think this one action is lifting her dress and she's going commando =D I'm trying more prompts in the next few.
Update: Nah wasn't it =D.
Everyday this kind of thing diluted the lora list.
I've classified this as AIO in my library. The examples helped.
im getting much better character consistancy at 0.6 strength but it does work nicely... hope this improves out of beta.. nice job
thanks
What does this LoRA do? It would be helpful if you could provide a clear description or sample video comparing the "on" and "off" states.
for me it doesnt did nothing if use above 0.20 it fuck the whole video quality and create same strange things, looks it work only 10% like 0.10 strenght, I didn't see anything worth using at all
This LoRA is absolute garbage, the video lost like 60% of its quality. MiniMax works way better without it.
I have also noticed quality loss on things like cloth and teeth at high lora values
I did a honest test and compared it on existing flows for which I used Mysticxxx and M3 unlocked, to be honest I found the quality way below that. Also maybe you can explain, as from what I understand a finetune extraction lora, works a bit as a delta on a basemodel with applied loras. Why do you then say it is trained? It seems me more an extraction than a training.
Your applying the SVD as a lora for most blocks it was 90% for a few it was 60%, average was still around 90% - That being said the alpha or rank may have been to high for this purpose
@Felldude Ok thanks for your response, I did some quick testing with combining your model with low weight (0.2 ish) with the above NSFW loras, from what I can now see it might be benefitial in smoother (and faster) motion.
Which finetune was this lora extracted from?
My own training