NOTE: as I am still learning best practices as I go and trying to improve, TRIGGERS ARE SLIGHTLY DIFFERENT BETWEEN VERSIONS!
V0.4
This is a minor re-do of v0.3 to make it better at halloween version specifically, since that was what the original bounty request was and it expires today (the day I am posting this version of lora). If you want to make original outfit roll, v0.3 might be better, or at least more flexible.
tags are largely the same, just "diamondhairclip" is back to "diamond hairclip" and her normal form dress tag is "classicrolldress," (I am guessing this is actually what screwed up the style lol, will keep experimenting)
also no longer a separate tag for her "cape", as that seemed to confuse the model too much.
Overall this one should be better optimized for reproducing her halloween form details, at least from the testing I have done so far...
Halloween form triggers:
megamanroll, halloweenroll, blonde hair, frilled gloves, twintails, pink legwear, pink ribbon, rollwand, pink dress, white frills, sidelocks, hair between eyes, bare shoulders, green eyes, pink ascot, diamond hairclip, magical girl, white shirt, hair ribbon,
Normal form triggers (note: if it keeps making the whole dress black instead of just sleeves, try "red classicrolldress," instead of just classicrolldress,):
megamanroll, classicrolldress, long hair, blonde hair, green eyes, ponytail, red legwear, green ribbon, sidelocks, long sleeves, white frills, hair ribbon, black sleeves,
Optional variables:
arm cannon, short sleeves, bare neck, teeth, upper teeth only, red classicrolldress,
You can try to mix and match outfit parts, but consistency of remaining elements will suffer.
This version is also reasonably flexible with the character as well, so you should be able to make custom characters with the same general feel as well to some extent. See example images for... well examples xD
V0.3
Since it has been so long since I mentioned I would update figured I would release something.
As mentioned since I could not not get a satisfactory all in one in time, I will be releasing two separate versions. This is the detail-focused version. (The style flexibility at the cost of details lora will be called 0.5 and released a bit later)
It is overcooked, so some style is sacrificed for understanding of outfit details, and you will need to use explicit tags and negatives for best results (open mouth/closed mouth, long/short sleeves, happy/sad face, teeth, upper teeth only, etc), but it now has a decent understanding of the various outfit parts, allowing you to mix and match them in creative ways as you please.
TAGS HAVE CHANGED!
You can try mix and match them for the most parts, but some are less compatible than others...
Main trigger:
megamanroll, blonde hair, sidelocks, hair between eyes, green eyes,
Outfits and appearance (add these to the main trigger)
Halloween outfit parts:
halloweenroll, twintails, frilled gloves, pink legwear, pink ribbon, pink dress, white frills, bare shoulders, pink ascot, diamondhairclip, rollcape, magical girl, white shirt, hair ribbon
Classic outfit:
rolldress, long hair, long sleeves, green eyes, ponytail, red legwear, green ribbon, white frills, hair ribbon, black sleeves, sleeve cuffs, turtleneck dress,
Other trained outfit and accessory options:
rollwand, arm cannon, blush, blue eyes, short sleeves, green bow, red bow, zipper, bare neck, hair bow, red dress,
Hope at least some of you find this one useful >.<
UPDATE2: still working on more experiments with re balancing the dataset as the 3 experimental versions I made so far seem like sidegrades more than upgrades. Will try one more time tomorrow evening, and If I can not fit it all in properly might just post 2 versions: one that is worse at details, but trained normally, and one that is slightly overtrained, but much more likely to get details right.
V0.1:
(update: I honestly did not expect this one will get so much attention, thanks and glad you liked it! ^^,... though I am kinda worried now that whatever I release after will not even qualify as an "upgrade" lol)
This is an experimental V0.1 alpha version, posted mainly for educational purposes to have something to link to when explaining stuff about loras... and to use as an example of what happens when you slightly mess up tagging lol.
It will produce the character and outfits, but it is not optimized and overtrained on some details so quality and compatibility is not guaranteed and some outfit parts are unreliable due to said tag error.
I will keep working on this one as I am using it for some settings experimentation to see what amount of options for modular outfits and detail I can get away with in a small-ish lora so expect to see updates from time to time.
Halloween version:
roll \(mega man\), halloweenroll, blonde hair, white gloves, twintails, pink knee boots, pink ribbon, pink dress, white skirt, sidelocks, hair between eyes, bare shoulders, green eyes, pink necktie, diamond hairclip, rollcape, android,
Default version:
roll \(mega man\), rolldress, long hair, blonde hair, green eyes, ponytail, red knee boots, green hair ribbon, sidelocks, long sleeves, android,
Optional parts (unreliable, but can get right with enough tries):
arm cannon, rollwand,
Description
V0.1:
This is an experimental V0.1 alpha version, posted mainly for educational purposes to have something to link to when explaining stuff about loras... and to use as an example of what happens when you slightly mess up tagging lol.
It will produce the character and outfits, but it is not optimized and overtrained on some details so quality and compatibility is not guaranteed and some outfit parts are unreliable due to said tag error.
I will keep working on this one as I am using it for some settings experimentation to see what amount of options for modular outfits and detail I can get away with in a small-ish lora so expect to see updates from time to time.
Halloween version:
roll \(mega man\), halloweenroll, blonde hair, white gloves, twintails, pink knee boots, pink ribbon, pink dress, white skirt, sidelocks, hair between eyes, bare shoulders, green eyes, pink necktie, diamond hairclip, rollcape, android,
Default version:
roll \(mega man\), rolldress, long hair, blonde hair, green eyes, ponytail, red knee boots, green hair ribbon, sidelocks, long sleeves, android,
Optional parts (unreliable, but can get right with enough tries):
arm cannon, rollwand,
FAQ
Comments (6)
Thanks for this, it's a really cute model! One thing I noticed (might just be model dependent?) though I couldn't get her to not have the open mouth happy face, even with happy, smile and mouth open/open mouth in negs, lol. My guess on why this happened, even if all images were properly tagged, if there's too much of one repeating concept in the image dataset, it will bleed into the lora if it's been cooked with too many repeats and not enough epochs
Thanks for your comment! Yes you are probably right, the dataset is fairly limited and I had to cook it a bit longer so that it understood the minor details/accessories (mainly was trying to have at least SOME chance of getting the wand and arm cannon right), and the expression might have gotten baked in as a sideffect. As I mentioned in description this is a V0.1 ALPHA version that I know has a few tag errors, I am working on a new version with the tag errors fixed and a slightly larger dataset that will hopefully fix these issues and it should be out sometime this weekend at the latest.
I made a note about the facial expression issue and will test that more before staring training of the next version (and add another image or 2 of her with her mouth closed maybe). In the meantime if you want to keep using the alpha version, I would recommend you use "closed mouth" and "frown" "sad" etc depending on what expression you want in the positive prompt. It will bleed sometimes by in my limited spot test the model usually responded to this. Also aside from "open mouth" can try putting "teeth" and "upper teeth only" in negative since that is usually associated with her open mouth.
This was originally just an experiment to work on over time and I had no idea when I released it that it will become my top model in under one day... I guess roll is even more popular than I thought xD
@S27 I'll try that, didn't think about teeth in negs, lol. Yes, Roll is quite popular, I think what helped boost it is the cute photo of her you used as the front pic. Also the size of the lora is a factor as well. Definitely looking forward to your future loras!
@Light7799 Thanks! The dataset tweaks for v0.5 of this lora are more or less done, just waiting for my gpu to finish cooking a few other things then I will see if I can improve this one.
Yeah I originally thought people would prefer bigger loras that can pack more detail into them so was working with larger network sizes... guess that was a mistake and I will have to go back and re-make all of them in a smaller 50mb-ish size instead of 100-200 at some point (or do people want even smaller? would be hard to fit multiple outfits and style at any good level of detail with that...).
Should have done some research before I started out .... live and learn i guess xD
@S27 Well, funny you mention that about big loras packing more.. That's kinda been proven it doesn't need to be large. I think the size of this lora is fine. Someone best explained it me putting it like this; a large lora is like a small object in a really large room, sure you can easily find the object and see it, but you don't need all that extra space. Afterall, if the model itself is 6gb and the lora in question is like 200mb, only 30 of those lora could fit into a model... yet there's thousands upon thousands of characters and concepts in a model.. so something doesn't add up having to have new characters and concepts added in to have to be so large. Of course, this is all through trial and error for me, I don't understand this technology on a scientific technical level so I could be wrong on some of these things.
@Light7799 (Assuming I understand this, please correct me if I am wrong)
Yes that is true that a lot of the time you simply end up with empty unused weights, as a typical character lora can be made with ~20 images and you will overcook long before you make use of a 200mb net with that (can probably get away with ~30mb in such a case). Thing is I kinda went overboard with those first loras I made and my datasets were more around 200+ screencaps per character, and the style ones were over 1k images for some, so my object from your analogy was a bit bigger lol.
This allowed me to transfer the art style, as well as small details in eyes and outfits that I had difficulty reproducing with a smaller lora. Obviously had I known then what I know now I probably would not have done that, as I realized most people here would value a compact and more flexible lora more than one that is more likely to get the details right, but is 4x larger and more likely to fight a style heavy base model for style control... and/or maybe I just overcooked mine or screwed up some other settings... could not find a good guide at the time that explained all the settings, their effects and side effects in detail, but I am guessing that with how the representation of those details was lower in the dataset, by the time they got trained some over-training occurred in other areas.
I will be remaking my earlier loras in the ~50mb size range when I get the time to make them more convenient to use as the details lost would probably be only noticed by die-hard perfectionists anyway...
In the future I will be sticking with ~50mb for characters (less if single outfit) and ~100mb per normal style probably, with rare exceptions where I make use of my older datasets to make all-in-ones and heavy multi style loras with 1k+ datasets. Not sure about concepts, but I guess that is somewhere in between style and character in terms of optimal size?
Anyways thank you for taking the time to comment and give suggestions! Wish I ran into you sooner... would have saved me hours of work making excessively large datasets lol
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.