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 (24)
Always interesting to see what you're cooking, keep up the good work!
Thanks
If I wasn't having issues uploading lately I'd have a batch done for this model. I keep getting mass failed to upload errors since a few days ago.
Once that is sorted out I'll see what I can do.
The site itself has had huge issues with the gallery for a month
looks interesting, ill give it a go
im getting great results even with my complex workflow. thats a great sign for this model!
The thumbsdown for complex workflows, must be comfy haters lol
Wow...
this looks kinda funky but I'm interested to see how it blends with other models, it looks like it's got some magic in it.
I get it, but this model's creator trains his checkpoints and puts in a lot of time and effort. I can see how a comment like this would be more discouraging than positive. "Your efforts are noted, but it falls short of being great on it's own". BUT, that is just me. I can't speak to what this creator feels about it.
I just admire his creative drive and his models certainly reflect his love for the medium. Maybe let him, and others visiting the page, know what make this model specifically great.
ironically_irascible Thanks for the encouragement towards positive. I don't take the your model won't do this or sucks as this as a negative by default, assuming it is accurate. While billions of parameters sounds like a lot it is really not if you consider you have a 47k vocabulary and if you could equally map that distribution it would be 47k to the power of 75 which is incomprehensible.
fp32 🙏
likely will be released at some point
any plans to revist PERFECT -32 bit projection? or add to the 3 existing versions? i used Body and Pose/Layout often, it works well even with alot of the conventional XL checkpoints
With what focus?
@Felldude I dunno... the focus of a guy who made 3 models that besides "Fp32 clip" were unrelated but made them part of the same model/series anyway and then stopped
@Legend_of_Chavo_G I am not working on anything other then large datasets in the TB range. And determining if any meaningful large scale training can be done in these larger models without a 80GB card or data center setup. Regarding the related/unrelated I group models with 90% of shared training data though the 10% change can effect allot
This is the most realistic looking SDXL image so far i think. really fast generation as well. Wow!! But prompt understanding is not that good. Hope will improve in next update. thanks
Here's an Example:
prompt-
A photo taken with an iPhone using portrait mode. The photo should look like it was taken by an amateur photographer. The photo is of a 23 year old woman. She has an Elevated Indie Aesthetic. She has unconvential good looks. The woman is wearing a bustier and a short trumpet skirt. She is average height. She has realistic skin with a minor imperfections.
Generated image: A woman is holding iphone and trumpet (music instrument) at the same time, white hair, 70 age old woman, wearing bra, smiling.
I did notice with it being so far finetuned toward human and nature it lacks some understanding of other prompts.
Felldude I think you can improve it (prompt understand). I know will be No.1 Best Checkpoint model 👍🏻
Ah on that the LLM's do not use 20yo or 20 year old. Milf was added into the LLM context but not age triggers
Felldude I dont know, is there any xl model, which knows trumpet skirt out of box.
The prompt is beyond shit too.
Felldude Much appreciated for your concern. I tried to fix it loras for clothing but she wears pants instead and hold music trumpet. There's a casual skirt called "short trumpet skirt". But i think it struggles to understand this prompt. But anyways, i like the realistic color of this checkpoint, quality and speed generation is much faster than other sdxl models. I can use other Loras to fix around. Thank you very much.