Regular lora version update: Lora version trained with other hyperparams, should perform better, even with innate artist OC's prompted.
Old info About the DoRA part: you need to apply this commit with forge or update to the latest commit of a1111 to use the correct implementation of dora, otherwise it will work just like a regular locon.
Lora for this artist is redundant, since pony knows the style without it. It was trained with two purposes, first to remove signature from gens, second to get more animefied look of gens with it. List with all tags can be found here.
Description
DoRA version
FAQ
Comments (14)
Is there a page where I can read more about DoRA? It's new for me
You can read the paper, if you really want to dive into it https://arxiv.org/pdf/2402.09353.pdf but for simplicity its just not well optimized right now to train, but theoretically should yield better quality after training this way
@bakariso Thanks!
@bakariso It does yield better reproduction of styles and some concepts do take better with DoRA, the issue is the 30 to 60% increased training time with same hardware, but trust me on this, it IS better.
@bluvoll Its not that drastical of a difference between properly trained locon and dora in quality from what I've tested. Speed slowdown is because kohaku fucked something up, it's dropped even for locon with latest update
@bakariso You´re absolutely right, but that 1 to 2% is sometimes worth dealing with w/e Kohaku did lmao
@bluvoll It's not 1-2%, maybe 5-10, if you willing to wait, surely dora is preferrably, not sure if dora should trains even slower, but it is, despite of the bug with locon speed
@bakariso From my anecdotal testing, at best is a ~~8% but varies with dataset, small datasets(up to 70ish images) enjoyed up to 5% but most of the time wasn't that noticeable , for bigger (1.3k images or more ) I did see a noticeable improvement but the overall time used for the DoRA made me cry, plus we rarely use such big datasets for LoRA.
@bluvoll What are your measurements based on, just an empirical feeling or really measured statistically?
@bakariso Statiscally tested using my Style datasets which range from 70images, Hiroshi, to 1.3k images Asanagi, and then blind testing with some folks regarding which one they preferred, and most of the time they picked DoRA over LoRA on large datasets, but didn't have strong preferences when the dataset was small, testing took some 2 weeks even with one A100 as training time went up 30% on that card after maxing out possible batch size, in comparison A6000 increased by some 40% and ~~50% on my 3090.
@bluvoll Hmm, have you tried not maxing batch size for small datasets? Its not very good idea to go for something bigger than 2 for 70-100 pics, but absolutely fine and preferrable for 1300
@bakariso I tested all the way down to 1, and changes were minimal as I tried styles, not characters or concepts, so those are outside the scope of my testing, sadly.
@bluvoll Thats weird, because batch is one of the main settings that will affect style comprehension. Have you collected all those tests somewhere to look and compare?
@bakariso Just for personal use as I use a heavily customized sd-scripts.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.