This is the XL lora version of the Alkemanubis (Alke) Style, based in the model Pony XL V6 and Illustrious01. This should works with any merge or derivated models.
Illustrious01 v1
I know that Illustrious already have this style by default, but actually it looks pretty desaturated and the characters looks like deformed dolls. I imagine that it's due the existence of old works and sketches in the dataset. But with this lora you can get more vibrant results. Additionally, I've updated the dataset and pruned it a little.
Weight: .8 - 1
Pony6 v3: LoKr
After some tests, I don't want to know nothing about DoRA's for some months xD, they are rlly stranges and gens rlly weird results. Now I have to re-train the DoRA's that I published. Sorry for it -.-. I take advantage for change for a new tagger and prune some images from the dataset.
Weight: .8
Pony6 v2 (DoRA)
Maybe you'll need to update your Auto1111 to version 1.9.0 or take other steps to make it work in Forge/Confy.
This is my very first attempt at creating a Dora, so it might be a bit undercooked. What are my thoughts about it?
I'm pleased with how much I can increase the weight without changing the image, just the style.
It seems to require a higher learning rate than LoKr or LoRa. I'll experiment with other datasets to test this theory, but according to the paper, this type of Lora is supposed to train faster... I don't know xD.
By the moment, I still prefer the LoKrs, but I'll try again maybe other 2 or 3 times before abandon it.
Weight: 1 - 1.6
Pony6 v1
Trigger: by alkemanubis
Weight: .8 - 1
Description
FAQ
Comments (6)
was dora that bad, sorry haha
Haha, well, in my case, I conducted around 6 tests with different training data and variations in parameters. I obtained results such as undercooked models (like this one in version 2, which was trained for around 3 hours and the style is not noticeable in some prompts like the 2b image), models that produced really strange results like the Ookami, which became messy if you added "monochrome" to the neg prompt (this didn't happen with Loras/Lycoris; in fact, in Lycoris, it helps to get more vivid colors). Additionally, that one had really hard problems with hands and bodies in general -.-. I wasted my time training another 3 unreleased models because they didn't capture the style or required weight to be increased to 1.8, resulting in a lot of deformations (those took around 12, 8, and 4 hours to train each one, with the same terrible result using parameters from other creators here). The only one that looks decent is the Porforever Dora, but this one is pretty weird; if you prompt for simple poses, it works pretty well, but if you ask for a pose that is not standing, you can get results with characters holding strange things or even looking at the same place. The results seem like a SD1.5 Lora, what the heck... Additionally, this requires more resources for training and generating images, isn't compatible with tools like merging between Loras or to a base model, and has a pretty immature compatibility with some generators...
@Konoko Also some people, me included, have DoRAs shoot up in vram usage when using with comfyui, even on last commit with support suposed to be working already. A1111 seems to work fine though.
@hhks Well is time to wait until DoRAs gets a more stable and compatible status in the SD enviroment (because it seems that it works pretty well in LLMs)
Sorry. I feel responsible for sending you down the DoRa rabbit hole by asking if you knew anything about it. You seemed to learn some useful info about its weaknesses/strengths at least?
I take my own (bad v:) decisions, was more for curiousity. In the paper, DoRA's looked rlly interesting. But maybe, they are more delicated than a regular Lora. I'm not sure if I'll try them again in a future, but I'll keep looking for improve my Loras c: