Complete and total rework of this new version, Free, this will prevent some assholes from getting paid on Patreon
PS :
I won't be giving charity to get buzz. If you appreciate this Lora and my work, I'll let you express your support by commenting or sharing an image you created using this Lora.
Thank you ;)
Using the checkpoint model: waiNSFWIllustrious_V140.safetensors, Lora strength recommended 07 to 1.
Trigger : Gachiakuta_style, masterpiece, best quality, official art, bold lineart, gritty, dramatic lighting, dim lighting, shadow on face, hatching (texture), ultra-detailed, highres, hyper-detailed, fine details, 8K, Vivid color, intricate details, absurdre.
Description
Complete and total rework of this new version, Free, this will prevent some assholes from getting paid on PatreonFAQ
Comments (11)
Bruh, this is the first time I'm seeing a 2GB LoRA. At this point you might as well have done a full-on finetune. You don't need this much for a style. Even 200MB is sometimes overkill
So :) to capture the style completely and integrally, you can forget the 200 MB, I did the lora in several steps in 4h00 training sessions: 16/8, 32/16, 64/32, 92/48, 128/64, 256/128 and only this different way that I chose allowed me to obtain the faithful style so thank you for the advice ;), but you have 200 MB, would have provided only 1/10th of the rendering.
@AugusWrath I find that very weird... Did you change anything else in your training approach, besides the dim/alpha? Barring very specific artificial corner cases, SVD provides diminishing returns as the dimensionality approaches the original, so to have such a drastic gap in quality between a very heavy and normal-sized LoRAs is surprising to say the least. 200 MB wouldn't be 1/10th of the capacity, more like 95 to 99 percent.
To put it another way, consider the volume of information: 6 GB is enough for Illustrious to know with good precision thousands of characters and styles, not to mention concept and composition understanding. This is significantly better than, say, Pony or base SDXL, but not due to any increase in the number of parameters -- only because of better (and possibly longer) training and a better-tagged dataset. The vast majority of the LoRAs on site are under 250MB, and fare well enough that nobody (almost nobody?) decides to retrain them at a higher dimensionality. This includes very unique styles, and styles + character packs, and complex concepts, and... everything, really. Even realism (I mean actual realism, as in "indistinguishable from a photo of a real person", not 2.5d or 3d or whatever), something that requires probably the most precision as far as styles go, has been done convincingly in 400MB LoRAs.
@Abel1502
Listen :) , I have 8TB of Lora creations on SD1.5 and SDXL: Pony and Illustrious.
I've been creating Lora and Checkpoint for many years.
This Lora is 1.7GB, which is nothing at all, no one's complaining except you.
Everyone appreciates Lora, so if your graphics card doesn't have the capacity to run Lora, it's not my fault. From now on, I'm not going to get into speeches and discussions that won't add anything. It's precisely because there's always someone who reports their knowledge that I don't normally post my Lora or Checkpoint.
I wanted to please the community, don't make my choice to please the community turn into the opposite, and I remove Lora.
@Abel1502
and I will end up telling you, if La lora does not suit you :) create your own in 200MB or 400MB :)
@AugusWrath I'm not complaining or blaming you. Sorry if I came out that way. I find your response unreasonably rude. What I was saying was that your choice was surprising to me, and that I suspect there might be an easy way to optimize your approach in the future. I'm not forcing you to do anything, and I agree that any lora is better than no lora.
@AugusWrath If I ever need a Gachiakuta style lora, there already are several available on civitai that appear sufficient to me, but in general yes, that is what I would do
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.