FinalCut (SDXL)
Onsite Recommend Settings, 16 Steps, CFG 5, DPM++ 2M
High Rez Fix, Image to Image at 0.7-0.8 remove fixed seed.
Do not use 20yo or 20 year old as Llama was used for image captioning. MILF was added. Use of Old is like using GILF
13GB model has better general prompt adherence, while the 8GB model is focused on skin detail and anatomy text prompting.
Full FP32 Training
8GB model Mixed BF16/FP16/FP32 storage ensures highest accuracy while reducing model size.
All cover example images shown natively at 1024x1536, however this model will benefit greatly from 1280x1920 image to image up-scaling at 0.7 and has been trained for such.
Description
FAQ
Comments (30)
will the 13gb model ruin me? on like 8gb vram rn
Yes
Felldude dang. dont wanna bake this potato before upgrades. doing great work btw!
dukesups It will run fine if your forcing the CLIP into FP32 (CPU) and downcasting the unet to BF16 or FP16
Full FP32 would increase iteration time by 5x-10x though
What do you mean you put 50k?
50k buzz has the model in the top 10 for onsite use next week but could easily change
Oof ok I thought he spent 50k$ for training the model lol
Better than lustify. My new go-to checkpoint.
my image comes out grainy
and sometimes with bad face.
any suggestion to get quality like your sample pics.
The model was intended for use at 1280x1536 or 1024x1536 natively - or 832x1236 with image to image 1536x1920 at 0.7
thank you <3
nice
First of all, thanks for this model! I'm using the pruned fp32 model, and I have to say that I've tried almost a hundred models for my advanced ComfyUI workflow, including all the big names, of course. In terms of image quality, this model produced higher quality than all of them. Many people misuse or misunderstand SDXL. If you use a well-structured workflow that relies on upscaling with hiresfix, SDXL can achieve FLUX quality. Since it is essential to use higher tile resolutions above a certain resolution during upscaling, it is important how confidently the model handles higher resolutions. This model is very good at this and it handles higher resolutions excellently. Thanks to its fp32 precision, is capable of achieving amazing detail. The creator of this model has created something great! From now on, this is my number one model. To the creator: man, you've created something huge, I can only hope you continue your great work with this, because if you manage to work a little more on prompt adherence, this model will be unbeatable! It has so much potential that you can knock all the big names off their thrones with it!
Thank You - It seems that average AI is shifting toward smaller portrait size photos with "high accuracy" rather then HD or UHD photos
Hey, could you share such a good workflow? Seem like a gem
this model is awesome no doubt about that, but when you compare other models do you also use them in FP32 ? because comparing FP16 to FP32 is kinda biased from the beginning.
@napsumfur225 To my knowledge base sdxl is the only other fp32 fine tune - in my own comparison it is far better at text and generalization but not nsfw as that was all blurred by the training script
Hi, are you able to share your workflow? I'm still trying new things here
What upscaling method do you mean with 0.7 denoise?
Image to image
@Felldude I wonder, if there something about your workflow that is different? For me, using img2img in Forge UI, I start to get slight deviations as soon as I go higher than 0.3 denoise.
0.7 tends to completely rearrange the picture. Can you elaborate a little more on how you do it? Great checkpoint by the way - I get fantastic results just using 1024x1536, DPM++ 2M, Karras, 20 steps, CFG 5 in txt2img.
@nroonarij129 slight deviations at that level I would expect, however I would say 70-80% of the images keep the general layout and feel either using the same prompt or a blank prompt. I do think auto1111/forge might be making the difference with the way it handles prompting
But can it do really fat girls like EpicRealism??? The question on all of our mindssssssssssssssssssssssss
I manually captioned some breast size triggers but not extreme body shape triggers
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.