🎵 Like a virgin, training a LORA for the very first time.... like a vir-ir-ir-ir-gin, tuning trainers, going out of my mind.... 🎵
Not my first time, but it can feel like that when you re-caption for the fifth time and only then prune half the set because you realize that awesome clips don't always make for stable training material. uughhh
The front doggy LORA is fun, but I found it annoying that we couldn't see much of what was going on behind. Initially I clipped a set that gave higher angle views and some slightly off center, but then it quickly became NOT front-doggy. Since we already have decent front-doggy, I decide to angle slightly to the side, and this is the result (so far).
I will work on and improve Front Doggy again, as I have a lot of great clips, but that will have to wait.
As for THIS concept, I've already re-captioned and am training again right now but this version ain't half bad!
I'll add a wildcard prompt and training details when I get a version of this that I consider closer to final. I assume y'all appreciate these pre-release LORAs while waiting for the more stable ones.
Also, sometimes training runs give me a worse result, so who knows, maybe this is as good as it gets and a better prompt will unlock the magic with less seed hunting.
Stay tuned...
Training
I can't stop improving my knowledge, tearing down old bad habits, making the same mistakes again and again, and relearning to walk, before I run, but the core principles have not changed, so check out my training article.
Disclaimer
Prompt responsibly.
Description
FAQ
Comments (13)
thanks for all the work you put into these, I can't wait to give this a test drive!
Why Hunyuan's nsfw creators, still haven't migrated to wan🫤.
Framepack is where it's at.
Open to it, but... I tried WAN once, it was slower and seemed less stable with worse NSFW performance. Was that just a bad first impression? Fact is all my sets are intact so i should be able to just reTrain for WAN at some point, just, I don't use WAN and this stuff is extremely time consuming. Testing LORAs and prepping showcase materials alone takes hours per LORA, then if testing doesn't go well I clean my set more and retrain. Each LORA takes 1-3 days minimum using this strategy, not including initial dataset prep.
Aside from all the awesome comments, I'm not sure why I do this :P
@az420 Have you tried wan fusionX or wan Causvid Lora? Life is better now for wan. Wan nsfw isn't that bad and rarely requires other loras to work well. It's just hard to go back to Hunyuan, after that easy wan👌🏻.
@Mikebobby Apologies for my ignorance, but... how is WAN better than HunyuanVideo? You say it's hard to go back to Hunyuan... in what way? This is not a challenge, I fully expect models to supplant each other eventually, so I am curious :)
so I tried fusionX yesterday in a decent workflow. Results were good, but.... the workflow I use gives me a preview of the video at low rez in only about 10 seconds. I guess I remember before FastHunyuanVideo and Teacache HunyuanVideo was that slow too.
WAN did do a really good job, even without a LORA, but... it took like two minutes to render an 81 frame video, I just don't have that patience :P
@az420 In case you were wondering, the reason I'm suggesting framepack is not so much the quality, but rather the length of video you can generate on a low spec card. Easily more than a minute of video. Also, Framepack Studio lets you specify different prompts at different timestamps so you have a lot of control through the video generation.
t2v hunyuan > wan
i2v wan > hunyuan
I've been using your LoRA with Framepack's i2v setup, and while it's really cool, I noticed that compared to the 'front doggy' version, this one changes the face a lot—it ends up looking like a completely different person sometimes! Still, I really appreciate your work. Thank you so much, and I'm looking forward to seeing more of your creations!
Two things: The version currently available is not the best, it was good enough to release as a sort of beta version but yeah, not final. With that said, I hear that if you want a consistent face, it helps a lot to prompt for it pretty carefully, like "thin eyebrows, blue eyes, lipstick, cute nose" or whatever, to try and get the face coming from the same weights.
You should change the main showcase video from 82326694 to 82327223. It describes better and won't being confused with "Front Doggystyle" ... just imo
Didn't tried yet but sure Thanks already !
I spent hours yesterday test a retrained version, it seemed to lock into ONE model's face and her exaggerated head movements which kinda ruined it. I pruned her, and another video I found was over represented, but then ended up with 5 clips, which is barely enough. So... I've mixed in a few more clips that aren't perfect, but may give us better variety. I've also blurred some of the faces so fingers crossed that this new run does better.
Otherwise, the current version really isn't THAT bad, and since it's kind of a unique angle, it could end up being the only version... When dataset prep takes too long it really kills the vibe... gotta stay motivated, you know?
stay tuned!