Consistent Face (Wan Video 14B t2v)
Trained on 3D video to enforce a consistent face across T2V or I2V
Low Noise Training
Single Shot Training not jump cut.
Training was a single shot of a high detail (Full rendered frames) 3D female face to enforce consistency across a 5s (300 image) video.
Description
FAQ
Comments (49)
This is a character lora? or help characters face?
The training data was 3D video fully rendered frames to help with facial feature consistency
@Felldude Consistency of whose face? This one girl? Any face? Does it keep faces better over scene transitions? From your description alone I can't tell what this LoRA does, and your comment just repeats the description.
@boz255 I don't have the ability to generate the thousands of test images I would normally do for a lora. I can say based on a few images it learned the 3D source face geometry beyond that you will need to test yourself
@Felldude sorry man, thank you for releasing something for free, I think there is just a miscommunication: what is the purpose of this lora is all we want to know. what was the motivation to make it? Does it help keep the face the same during the video? Or is it trained for some other purpose, if yes, which? I don't think anyone understands what you mean by 3D source face geometry
@Radyschen A 3D character face as in CGI, it is a likeness but not of a real person
@Felldude So it generates a specific 3D character's face? Then why did you name it "consistent face"?
@Radyschen Yeah, his explanation is incredibly vague and seemingly like it was written by someone with dementia.
I think the obvious question is this: does it keep ANY face consistent or just the one pictured here on the lora download page? If the former - it would be heaven-sent for us all, if the latter, not interested (sorry).
@dcham2310 In 20 or so runs the face was consistent, however that does not mean that it would be that for 100% of the generations.
@DaddyWolfgang I think maybe I have dementia now after reading this thread.
@Felldude Sorry - still not answering the question (is that deliberate??) ... do you mean ANY face?
The OP's response seems like a prank 😂
I feel like i'm at a congressional hearing from the way you're talking around an incredibly easy question
Is this for HN, LN, both? Can you add some more description please?
It was trained on low noise
Great Lora👍👍
I love it
Show some results lol
Maybe you can help answer everyone's question that OP can't seem to understand. Is this a character lora of the OP image or does it help keep other peoples face consistent.
Tell us what it does, wise one!
Write me a recipe for blueberry pancakes
I think what OP means is a single shot that is consistent in face quality. No weird morphing or glitching. Probably training on high res closeups with no cuts
Here is a side by side comparison with a I2V workflow. Watch it and decide for yourself: https://streamable.com/sk49aq
Instead of bitching around, just download it and check if you have any use for it. There’s work put into this. It’s unbelievable how people who contribute nothing can talk so much against those who actually try to do something.
Lol, so it does nothing? Got it.
@GayLizardSpy The major difference I can see is that using the LoRA makes the character keep the mouth closed a little more which is more like the image input. I don't know how it will affect videos with extreme facial expressions and/or perspective and zoom changes. Also it might have different impact in T2V.
Appreciate your work. I would have thought more users would be aware of the identity consistency issues, but maybe it's more of a problem with I2V and their confusion is because this lora details say it was trained for the T2V checkpoint. Either way it's still an appropriate problem to tackle, particularly when using multiple loras which often results inconsistent subject identity for my I2V generations. If you haven't seen it, check out the phantom framework, used in the Wan 2.1 Phantom model.
It looks worse with the LoRA, so I'm not really sure what you're trying to prove here?
@meropewantsstls I am not trying to prove anything. I stumbled upon this lora and instead of spamming the comments with bullshit, I just decided to do something useful. As I said, watch it and decide for yourself.
I appreciate you doing this test and sharing it. I often wanted to do the same thing, but never thought to put it up on streamable. I'll do that for more of these loras where people keep asking questions. But yeah, I think any differences about the mouth being slightly more open or whatever isn't in the realm of face consistency. The only question is, "does it look more like her with the lora when she opens her mouth", instead of "does she ever open her mouth slightly when using the lora". I mean, video is supposed to make them look alive, and human faces do stuff like part lips slightly sometimes.
No one is "bitching", people are confused by the overly vague description.
Thanks for sharing the stream. I can't see much of a difference but I am impressed that you managed to get a "No Lora" video that looked so much like the original. My efforts always end up with only a slight resemblance at best (a distant cousin, maybe). Care to share your process? *** [EDIT] *** Not sure how I missed it but I now see that this Lora is for T2V -- that explains a lot. I was imagining an i2v with face consistency.
@WatchemAI Yep, that was me thinking it was for i2v.
@dcham2310 I tried the lora on I2V intentionally because there we have a face to begin with. Maybe I should do the test in T2V with a character lora and see if it makes any difference when the character moves out and back into the frame. When I think about it in detail, it should not affect a character lora that much because the model already knows how the character looks from the lora. If at all, it could help when generating a video with the prompt "a woman" without using a specific character lora. I might check it out some time
@fenasikerim I wonder if trying to go beyond 81 frames is a better test to see if it can keep better coherency where we know coherency breaks down with standard videos.
@Jellai I created the sample videos with 257 frames intentionally. I might test it on T2V without a character lora. Lets see if I can catch the same character using the same seed.
No one should have to download and test shit. Users are required to provide a description and examples up front so people aren't wasting their time, bandwidth, HD space, and GPU lifespan and random shit some schizo fucking porn addict threw up on the internet.
@mobdik17378 Yeah sure spoiled gen z prick. Expecting everything on a silver plate. How can people even dare not to generate the videos for you too. Oh wait why don't people just jerk you off directly so you don't hurt your hand? Take it or leave it. You ain't paying shit so you have the right to demand exactly shit. Now get lost.
@fenasikerim Doing the bare minimum like everyone fucking else is not "everything on a silver plate". There's a reason your generation will die more reviled and hated than any other in our history, you fat worthless fucking morons
@mobdik17378 Let’s get something straight: the creator didn’t just sneeze out a LoRA in five seconds. They spent hours training it, tweaking it, testing it, paying with their either their own BUZZ coins, or using their own hardware, or maybe both, all so people could download it for free. That’s time, money, energy, and compute they didn’t owe anyone.
Meanwhile, you act like simply seeing a download button is some sort of personal attack on your bandwidth. Newsflash: no one dragged you to Civitai at gunpoint and forced you to download anything. It’s optional. Always has been.
What is required, though, is a minimum level of respect. When someone puts their work out there for others to use, without charging a cent , the normal human reaction is ‘thanks,’ not a meltdown about hypothetical GPU wear and imaginary storage crises.
Creators are the reason content exists in the first place. Without them, there’d be nothing for you to complain about. So maybe, just maybe, the people putting in the effort deserve better than being talked down to by someone contributing absolutely nothing except attitude.
If you don’t want to try the LoRA, fantastic. Move along. But pretending you’ve been wronged because someone shared free work they invested their own time and money into? That’s not criticism, no... that’s entitlement.
@OmegaWPN I see what you're saying as a lora creator myself, but all evidence points to this whole thing being a troll post, and people don't like being punked.
Really outstanding work! Perfect face result!!! Thank you!
very promising tech
Harms consistency aross scenes with transitions, which seems like the main effect you might want.
Good to know, was about to try using this for hard cuts.
Solid info. Too bad. That tool seems needed.
Jump cuts was exactly what it was trained to avoid with 3D video focusing on the face from different angles - For a video with jump cuts you would likely need to train on video with jump cuts that goes back to the same subject for the duration
Are you planning the different one with hard cut? If that is even possible to get? Thank you for your work!