Experimental model of Dan Mumford Style.
v1.0g :
This is an experiment to reach more coherence, its more tricky to use but it may get more truthful to the style.
v1.0 :
This is a regressed Recommended version of the model which has lower and softer weights. You can really go as 0.4 and 2.5 to get interesting results.
v0.9-rc-LoHa :
More iterations and training compared to v0.4
It can handle more weights and has much more style influence in it. Also I used refiner with same prompts in some of the examples.
You may crank up the weights like 1.5+ to get similar styling compared to v0.4, since LoHa is more flexible with weights and that way it can work much nicely with other styles.
v0.4-LoCon :
Added more training, it will handle shapes better but style might get extreme depends on the prompt and mix of other style tokens, so you can adjust the weight +/- around 0.5 to get more desired result.
Also SDXL + eularA and DPM++ samplers gives better generation with more steps like 600+ unlike SD1.5 (where the sampler stops impacting quality after ~60). So maybe it takes 6-10 minutes to generate an image but end results will be more defined, especially for background elements.
Support me here if you want: https://ko-fi.com/sensai3
Thank you.
Description
FAQ
Comments (4)
Hi, I come bearing bad news: Version 1g is severly undercooked. You can see the test in the post. I was able to crank up the strength of version 1g up to 20. That shouldn't be possible without breaking the image. I'd say version 1g can be scrapped, sorry. :(
Results: https://civitai.com/posts/542781
that's strictly possible because its a LoHa, not LoRa. Implementation of LoHa is different and it can go as much as 30 weights without breaking the image, depends on the training length and vector lengths. that's why LoHa can be used in style training where you could freely overdo the style if you want, or mix it with other LoHa/LoRa's.
And btw, that is why 1.0-LoHa version is recommended, because it's trained more like the optimal way.
@sensai3 I've never seen a loha/lyco with a recommended weight above 2. Guess, I have to do some testing there, but be that as it may, do you really want to claim 1g gives a better result than just using "Dan Mumford" as keyword at any strength? If you do, I'm gonna have to challenge you to a Dan Mumford battle, I think. :p
@OneViolentGentleman you can read about the implementation on LyCORIS repo, in my tests, more epochs you train a LoHa, more you could go high on the weights. that's the very best difference between LoRa and LoHa. for LoRa its the opposite, 'optimal epoch' and 'optimal LR' is the key for LoRa to success. but LoHa can be anything, that's obviously much fun to experiment.
not confident that 1g is giving more desired results over 1.0 but the thing is, I'm new to SDXL, the styles, latent space, token space all works differently than SD1.5, its not easy to test the coherence of the training without giving extensive testing to style keywords and combine them for better use.