馃馃巻NEW YEAR EDITION馃巻馃
I needed to reupload the version due to generator issues with uploading the fixed checkpoint that I updated. Same version that has been up for the past few days, so no need to re-download if you have it already. Just added some new pictures for fun.
I've been working on this model for the past couple of weeks and felt like I got it to a point where I'd be happy to release it. I've tried to create a model that has the colors and creativity of 1.5 FaeTastic and also tried to fix some issues that I feel SDXL has. I can't promise that it's perfect, but I feel like it's a step in the right direction.
I've trained several LoRAs that haven't been released for the purpose of merging them into the model, and I'll talk about the model genetics below. Hopefully openly talking about the way I merge and create models/LoRAs will open the door for others to learn as well. :)
Because I believe in the open-source spirit of SD, anyone is free to use this model or any of my LoRAs to merge or use as they see fit in an open-source sense, including selling any images or using them in personal projects. I would be extremely grateful to be credited for any merges or uses of my resources.
However, I am very opposed to having my model and resources I personally train and test and spend hundreds of hours on, be put up on third-party generation sites for their monetary gain where I receive nothing in return.
Model Genetics
Everything was merged with SuperMerger in A1111. All LoRAs were trained with Kohya.
Dreamshaper XL .5 with Rundiffusion XL .5
Added on top of that is the Fae Style SDXL LoRA
FaeTastic V1 SDXL .7 with ProtoVisionXL .3
Added SDXL High Details LoRA
Added SDXL Better Eyes LoRA
Trained a FaeTastic SDXL LoRA on high aesthetic, highly detailed, high resolution 1.5 personal generated images and merged in. This LoRA is pretty awesome, and I'll probably release it on its own as well.
Trained an SDXL LoRA on high-resolution, highly detailed images of realistic, drawn, and digital art of hands. These images were up close pictures of hands as well as images of people holding objects. This was merged in.
Recommended Resources
For those interested in training SDXL LoRAs I highly recommend this guide that Ally and I collaborated on together with a lot of extensive testing and training: https://civitai.notion.site/SDXL-1-0-Training-Overview-4fb03d5b4cbf482ebe13ffef7bcb8ead
If you ever have any questions or need help with my resources or anything Civitai related, I am always happy to assist, you can find me in the Civitai discord! <3
Description
FAQ
Comments (16)
Faetastic indeed!
this is a really faetastic model (I know, joke got burnt)
I know, it's a joke but I'm gonna take it for reals as I am a big fan of yours! <3
@Faeia聽Oh! didn't know, that means a lot to me, really!
Hm, an XL Checkpoint, with an excellent model pedigree specializing in Elves? Its almost as if the Universe is challenging me to run my 5 elves through this checkpoint to see what they look like.
Challenge Accepted.
love your philosophy and your creations. the community is a better place because of people like you.
thank you.
i cant run this for some reason, I get this error RuntimeError: shape '[10240, 1280]' is invalid for input of size 5265578
Really an awesome work fae !
Hi Fae, MASSIVE fan of your work and this is just incredible so Thank You.
Question: Not used SDXL much been playing in past day or so with this model (as been waiting for you to produce something tbh ;) ) But I notice on img-2-img (using some of my art) if I go above 40 or 50% for DeNoising I get very odd images that are just nothing but weird?
The images before this (say at 0.3 denoise) are just fabulous. I'm wondering if this is an SDXL 'thing'?
I'm using the "sdxl_vae.safetensors" and tested in both ClipSkip:1 & 2 but same issue?
Inside A1111 1.5.1
Any advice most welcome and keep up the great work. Must join your Patreon if you have one?
Sincerely,
S.
--------
OK, been playing around a lot on testing today and have discovered, that the issue of odd images only occurs if I leave the Sampling method on: DPM++2M Karras. If I switch to EulerA and set the denoising to higher than 50%, it seems to work ok? Strange. Will have to test other samplers . . .
Any input most welcome :)
S.
Hey there, I'm not having any issues, just tested it a bunch. I am using the dev branch of A1111 though.
https://i.imgur.com/ejwPb1y.jpg image that's been img2img with .80 denoising. If you can message me on Discord, I'm in the Civitai Discord or respond back here showing me the messed up images I can try to troubleshoot with you.
@Faeia聽OK, been playing around a lot on testing today and have discovered, that the issue of odd images only occurs if I leave the Sampling method on: DPM++2M Karras. If I switch to EulerA and set the denoising to higher than 50%, it seems to work ok? Strange. Will have to test other samplers . . .
(Just read your reply!) :) Will come back to you on Discord
Appreciate the response
@SkullyArt聽When img-2-img or upscaling, you have to be careful with # of steps and size and denoise all at the same time. If you crank it up too far, the AI will start to see small details that aren't meant to be there. A recent example from me was "forehead wrinkles" and I was using too many steps. It started to see them at the 4x upscale as "fingers" and started putting fingernails on the end of the furrowed brow. There are a lot of things that go into it generating "details" vs. "new artifacts." One size rarely fits all in generated images, so you should expect to fiddle with things a bit to dial it in properly for a given image. If you have a workflow that generates well most of the time, that's a win - but you should always anticipate going back to fiddle with some of these settings for certain images. Just my 2 cents ...