Peggy Bundy, the glorious MILF that everyone loves has made it to Hunyuan!
This character is from the hit TV show Married With Children and this lora captures her likeness excellently!
Mixes great with other loras, ideally strength between 0.9 - 1.1 (.95 at double blocks is great and has lots of movement).. Mix and match em all!
Works great with fast lora for low steps
Works best at 640x480, 512x768 etc...
This lora plays very nicely with most Loras out currently. Have fun!
USAGE:
Trigger word is not necessary in most cases, use it for extra emphasis when needed.
Double blocks at higher strength (sometimes 0.95 0 -> 1.1) or single blocks at lower strength.
This is a Lora trained of a real person so NSFW posting is not allowed.
Highly recommend using this workflow:
https://civarchive.com/models/1007385?modelVersionId=1378643
Description
v1.0 - Peggy Bundy from Married With Children
Trigger word is not necessary in most cases, but can be used to extra emphasis,
FAQ
Comments (16)
S-tier
Oh my gosh- a celeb LORA that actually looks like the person. And those animations are amazing.
What amuses me is the LORAs that utterly fail to capture the essential essence of the person, and when one points that out, the author hides the comment, as if that act saves their poor effort.
new to hunyuan but almost every Real Person lora i've tried is scarily good. maybe check your settings
some loras are trained using a different base model (like distilled) but then inference is done using fastvideo etc. In these cases you‘ll need to crank up the lora weight quite high, I find, before the likeness comes through - but it’s a balancing act to see how high the weight you input goes along with your workflow and mixture of other Loras.
Thanks for the kind words! I've learned a lot training my Loras and got the process pretty nailed down recently. My worst ones are Trump/Musk which ive fixed, im going to delete those and re-upload them tonight. I try to make sure they are consistent and actually capture the vibe/essence and not just trash like a lot of the ones here. Plus mine play well with other loras XD
Is it possible to train a Lora so that not every woman in the video turns into the person of the LoRa ? seems to happen with most person LoRas,maybe there is a way to avoid it in the videos but i havent figured out how yet.
It all comes down to how the model was trained with tagging. If the class of object as well as other tags describing appearance and settings etc are missing then the LORA will simply override the concept of 'a woman' with whatever LORA you use. I.E. all women will become the same.
It takes much more effort and trial/error to do this properly so most people just train with a trigger, good for single person vids but like you noticed it'll override everyone.
Essentially most of the loras out are trained in the quick and dirty method as they payoff is not worth the extra amount of work
@AstralFish Ah ok ,i understand, thanks for the explanation
@Pelvis I've read that that's an inherent issue/drawback of LoRAs (that all the women look the same)
Because a LoRA is a fine-tuned/trained "mini"-model it requires significantly less computational resources and memory compared to a WHOLE MODEL.
So in order for a LoRA to work it just modifies the original model’s weights rather than altering the entire model. I think that means it replaces the general concept of "a woman" with the specific person the LoRA is trained on? I'm 99% sure. If you want to get different women but don't have the computer power to train a whole simulated world of the women you want you have to use extra tools like regional prompting and face swamping and stuff that hasn't been made for Hunyuan yet.
@Shoescifer I see,thanks
Great work. Can you tell us about what the data set looks like ? is it videos or images ? what size and length and how long did it take to train ?
Thank you for the kind words!
- all of my loras are trained with images only (I've tried a few with videos but haven't gotten the process down yet as I do everything offline + locally).
- Training process takes ~1.5-2.5 hours depending on the quality of the data. I actively monitor training via tensorboard to view the loss graphs vs steps to determine when to stop.
@AstralFish - can you talk to me about your gear and settings? A 2 hour HY LoRA would be awesome.
@leisure_suit_larry yeah feel free to PM me ilm always happy to share my info (mainly just to take down those damn YouTubers who preach about "this is the only way to do this ps here's my patreon" lol.
You might want to change the keyword tag a bit, dynamic prompts uses double underscores ("__") for wildcard file names.
In the Automatic1111 version of dynamic prompts you could change the characters for that, BUT with the ComfyUI versiont by adieyal you can't, the github page hasn't hasn't been updated in 2 years, and I don't know if any of the more recent forks (last edited in 6 months) fix this.
I did look, but nothing shows in ComfyUI's settings section, and there doesn't seem to be a control panel like a few nodes have.
Nothing to worry about, dynamic prompts and wildcards nodes etc are used when the prompt starts with a double underscore. Having one in the middle of a word wind do anything.
I appreciate the feedback tho and thanks for much!