the irony xD

@LDWorksDervlex I was being ironic, bro! Good lora btw!

@fernando_souza but only people who were mad are white supremacists for Ariel having "wrong" skin color, despite her being, literally, a mermaid. SO not sure what's the point of irony since you're trying to make fun of people who were not complaining?

@Rin_chan I'm asian and at least in our country we don't see mermaids black. Mermaids are also popular in Asia, not just ariel, we also have tons of adaptations of mermaids from tv series to movies. Considering we have variety of color from pale, brown, black, we have always depicted mermaids as pale colored based on our mythology. Not to mention, from what we know in the animated movie, ariel was white. I wasn't surprised this black ariel wasn't popular in Asia and most of us believes its part of America's colorwashing.

@LDWorksDervlex Everyone knows that the good loras are at least 144mb, this one is only 36mb... Are you implying that they don't deserve more megabytes?

... im obv jk 

@LDWorksDervlex @LDWorksDavid Hi! Actual ML Engineer here to tell you this is some bullshit, got many training hours, manually implemented CLIP guidance, optimizer tuning, GAN implementation, SD contributions, and all that stuff. As usual, you know so little, people are all too happy to talk about what they do not understand as if it's expertise. "Do you know anything about unique tokens?". Give me a damned break.

LoRAs sitting on top of the preexisting checkpoint means that the thing you are trying to finetune is alignment, this is not Dreambooth training where you are trying to sidejack a hopefully mostly disentangled token because you want to wedge your foot in the sand without disrupting much.

> "Would I use a word like "Arielle" as trigger word, the model would also try to use it own learned word of Arielle."

Yes, the existing notion is part of the way there, and is enmeshed in the existing knowledge of a checkpoint, which you want to guide further, and finetune from above. You want to correct, steer, and upgrade the boat, not make a competing one crashing into the concept sideways.

But, instead you're actively working against the 150,000 A100 hours that it took to get Stable Diffusion's token understanding in alignment with CLIP's text embeddings on whatever concepts you are deciding to ignore by not using the actual words. You CAN make a LoRA for cats and use the phrase 'dog', and even force it to work, but that approach is stifling the quality and flexibility you can achieve, because you're working against alignment. This is effectively undoing the finetune.

Double Proof:

koyconi uses "steampunk", "bronzepunk", "cyberpunk" and these work fine to great. The AI suffix doesn't matter, they are tokenized left to right. There are "Samuel L Jackson", "Emilia Clarke", "Watercolor", and "Oil Painting" LORAs with their literal keywords and these work fine to great. There's a Norse Viking Warrior LORA with Norse, Viking, and Warrior keywords and these are all, gasp, fine. NOT n0r53vkngwr10r

Here's what happened: You saw people using "3mily" for their text embeddings to sidestep A1111's auto filename matching, and you just decided to keep running with that, and are now post-hoc rationalizing why that is a good idea in a sunk cost fallacy.

Next time, you investigate.

@32Bitshifter ok dude, chill a bit now? The one who came here with bad manners were neither my pal or me, in fact we shouldn't be fighting. If you don't see how you started this "fight" then you're blind,,, I'm not ML engineer at all, I'm just a professional 2D/3D artist, that's all. I have been training and playing with DD and SD for 10 months more or less so my expertise is little. Still I do what I can.

So... without knowing all you seems to know, I know that this afternoon I trained a "ryu" character two ways to test that. I trained it over SD 1.5 as LORA (alignment) and there were a strong difference than training with "ryu" (concept that AI knows) and "sfr1v" as token. And well, the first one threw me a mix of the already known concept and the new images which results in very very little of the new aesthetics, the most common factor was the face. However, the moment I trained it in a single new token the LORA magically got all the new aesthetics and the results improved vastly.

And since you put konyconi as proof I'm gonna proof my own diggings as proof too:
PureRef (gyazo.com)

Again, not ML Engineer but training with ryu word and not a new token for me or at least for ryu, is useless. I like it more the way "sfv" token represent it. Maybe is my taste?

If there is a big advantage on using non unique token then I will be using your way, period. But I look Ryu or any other characters already represented in AI and I find it mess with the training dragging less than it could of the input data. Hope this makes sense? Will try training it again and tell you how it goes.

And please, let's relax all a bit since we're here to enjoy.

EDIT: Did the re-training with ryu as token:
PureRef (gyazo.com)

All same seed/model/epoch trained with same everything. As you can see the changes minimal but still the old token fixed what AI is doing wrong in terms of colors (few things to be fair). It seems it has some problems with stances and I dont know why cause Rev itself does it "good" but the "ryu" token on the new LORA epoch is not getting it. So... changes like I said almost the same, few errors (which the old one didnt have) and problems with positions.

I'm talking with examples and not saying this is universal truth but what I see here with my eyes is telling me unique token (in this case only? weird) worked better than using the already known concept. And well, alignment (base model LORA train) was done in SD 1.5 not Rev or custom models.

I will be glad to talk more about this so we can clear all this thing but with respect and witthout harassing, OK?

I mean, maybe if you were paying the OP some money you would have a leg to stand on, otherwise your opinion is shit.

@robproctor83 nah, we don't have problems learning new methods or discussing things but with manners. Everything we are doing on the trainings is not random, is intended and Dervlex can tell you the amount of hours I have been putting for getting quality things since he met me -and the daily experimenting. (as I said I was tunning in DD and same for SD like 10 months ago and my IG is a real life fact of this) And same for Dervlex. So reading things like:

"You saw people using "3mily" for their text embeddings to sidestep A1111's auto filename matching, and you just decided to keep running with that"

Are false. But what is a bit annoying is the attitude. Other than that, I think all is clear. Cheers!