This SD 1.0XL checkpoint is based on my XL3 model
https://civarchive.com/models/118312/xl3-experimental-sd10-xl-fp16
So this XL3 is a merge between the refiner-model and the base model. Also I merged that offset-lora directly into XL 3. Then I added some art into XL3. This I added a lot of details to XL3. For the eye correction I used Perfect Eyes XL. The result was good but it felt a bit restrictive. So I merged a small percentage of NSFW into the mix..
Overall it was quite more complex than it sounds.. crashed a lot during merging...
Thanks to freek22 who helped me creating this. I highly recommend his new checkpoint FenrisXL. Check out his amazing work.
https://civarchive.com/user/freek22/models
This checkpoint is intended for creating SFW and NSFW Images, art and everything you can imagine. It's not intended for porn.
CaN Tips made a really good comparison video: (leave a like there)
Many images in my showcase are without using the refiner. native 1024x1024; no upscale. It works very well on DPM++ 2SA Karras @ 70 Steps . Euler a worked also for me. Feel free to experiment with every sampler :-). Use a low value for the refiner if you want to use it at all.
Tip: Use the SD-Upscaler or Ultimate SD Upscaler instead of the refiner. This gives for me the best results ( see the example pictures).
If you use Comfy UI.. Here is my workflow for Comfy UI https://civarchive.com/articles/1648/comfy-ui-workflow-xl6-hephaistos
You also can download the VAE and use it with the standard V2.10/V1.1/V1.0. (You don't need this for the 2.1_baked_VAE version) .
Here is the link:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main
Why FP16? Well it needs less VRAM to run and is better to train then FP32. This is very useful for those with less VRAM. Also this prevents an error when you want to run it in FP16 with A1111. Don't use the flag --no-half. To save even more VRAM set the flag --medvram or even --lowvram (this slows everything but alows you to render larger images).
Daedalus_7 created a really good guide regarding the best sampler for SD 1.0 XL
https://civarchive.com/articles/1501/best-sampler-for-sdxl
Flibertygibbet created a incredible guide for prompting and the use of the cfg scale with SD 1.0XL models.
https://civarchive.com/articles/1834
Thank you very much.
I merged those things into the model:
SDVN6-RealArtXL by StableDiffusionVN
https://civarchive.com/models/118114?modelVersionId=128795
first_nsfw_for_sdxl_v1 by sevenof9247
https://civarchive.com/models/118794/firstnsfwforsdxlv1
Perfect Eyes XL by Deizor
https://civarchive.com/models/118427/perfect-eyes-xl
V2.1:
freek22 helped me again training my model with some high-res pictures also some parts of his project FenrisXL made it into the model.
https://civarchive.com/models/122793/fenrisxl
Thanks for all your work.
Give me some feedback when you are pleased with the results. Also I will link my other projects down below ;-)
Also if you want to support me you can do this here.
Description
First Version
FAQ
Comments (21)
Mmmh not sure about the merge of base model and refiner model as the refiner is an image to image checkpoint for "refining"/denoising some kind of noisy raw image generated by the base model.
To be honest here I don't know what it actually does. I forcefully injected it as Lora. I tried it a couple of times and with some weird settings it somehow merged together with another Lora. Then I merged the result with the Offset Lora into the base model. I know that it has some effect XL1 and XL2 loved to crash way before 1024x1024 and XL9 (which might looks even better than XL6) just uses more VRAM and loves to crash randomly. So now to the effects of this forceful injection: it's might be subjective but I think it is slightly more detailed. The effect is not very strong here.. more in small details like the eyes or minor stuff. In the showcase for XL3 I have a adventure cat compared with the base model. Overall I don't have the expertise to finally give a verdict about this. I tend to use extensive trial and error methods to improve my work. For XL6 I created over 30 different checkpoint's and merged the best ones together.. again and again and again... So what I can say is that XL6 is a great checkpoint. It even suprised me how good it actually is. But as always there is a lot to improve... It's a rabbit hole I am glad following in ;-)
How did you manage to merge lora into the model, the option doesn't seem to work in Kohya?
@Afroman4peace聽Thx, I will try it ;)
here is the guide from freek22 https://civitai.com/articles/1540/how-to-merge-loras-in-stable-diffusion-xl-10-models
@Afroman4peace I had originally gone through some trial and error with achieving this myself. I originally tried SuperMerger in a1111 and that seemed like it will work (as you said), but I encounter vram issues when merging checkpoints -> lora with SuperMerger. However I can merge checkpoint -> checkpoint (sdxl base) just fine and do so still there for flexibility.
I then tried Kohya ui with no success, didn't seem to support sdxl base yet for lora merging.
I did find good success with merging checkpoints -> lora or lora -> lora in Comfy UI however. Fast and seems reliable? Thought to share as that might be a better option for anyone if they are dealing with a lot of crashes in SuperMerger/a1111 etc. currently 馃憤
@Ocean3聽I will test comfyUI this week. I have managed to use my SD 1.5 checkpoint as refiner and upscaler. For the VRam issues. freek 22 told me to use Supermerger. First I encountered a lot of crashes using his initial method. It was like a gamble for me. I have 12 GB VRam. Then I found the solution why it crashed. Every time I had loaded a tiny checkpoint it worked. Freek 22 refined this method again when he took a Lora into the model folder and loaded it as checkpoint. Then it worked nearly every time.
@Afroman4peace聽Thanks, I don't believe I tried any pruned checkpoint as things were a bit early on. In the past I've tried to merge LoRa as a model from the models folder with SD1.5 and had some issues. I could maybe look at trying something like that again.
4me its always generate asian faces, why? even with neg promps and pos for european faces...
I had some european faces. what prompts did you use?
I will test this out though
@Afroman4peace聽ok its fine atm, just did in neg (asian:1.3) and no asian anymore on renders, good
@saytomepls聽Thanks for your tip. I may will write this in the description
This is a common "bug" of all xdxl1.0 Model just add like mentioned Asian in neg prompt
So far, without a doubt, it is one of the best models I have tried for SDXL. Can I ask you how you create the checkpoint? Do you use any special program, do you have any tutorial that you can recommend? I wish I could do my own tests.
freek22 gave me the idea how to merge Loras into 1.0XL. I refined this hack. He will write a guide about this. I tend to use extensive trial and error methods to improve my work. For XL6 I created over 30 different checkpoint's and merged the best ones together.. again and again and again...
This model creates some freaky shit 馃槄
seems you can just dump a thousand words and it will spew out quality images with scary results
thanks :-). Then I reached my goal.. Thrust me I have some more freaky stuff just waiting to be published...
Looks really promosing but I tried to use it in A11 (1.6.4) and only get an error while deserializing header: MetadataIncompleteBuffer
I have found this about it. https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10199
Are you using the civitai helper? If so download it directly from civitai. But I will look into it when this isn't the issue.