This is a model I have been working on and then started merging a Lora Hands in (4-5 times) It follows prompts well but I really want people to test the hands. Only thing I dont like about this model are freckles. If you like it, take it and name it what you want. Ill just leave it up for people but is just a test for me. Im still working on it but I dont think its bad where it is right now.
Description
FAQ
Comments (13)
Thank for the hand images, it was just a test and was worried it wouldn't make a difference
I'm running thousands of tests to find out what it can and can't do; it's blowing my mind, it's amazing.
@Tarendax thank you
@Tarendax Ive seen many of your images, very good at righting prompts and you have an interesting imagination.
@jrrytng467 And I don't publish everything I think of or try. This model is giving me too much freedom; I haven't found its limits yet.
@Tarendax let me ask you this. Do you think it comes from the prompts? Im new to this but I read peoples prompts and I think even if you have a bad model that you can still create great images?
@jrrytng467 Well, I don't know. After using many models, I've only learned that each model is a different world.
You might start seeing images similar to past images, I'm just using the same prompts on a model I'm working on just to compare. Your getting a look into whats coming is all
@jrrytng467 This one is quite nice as well! Think I might like this even better than test_V1 (which I like better than Prompt4Pleasure v3.0 iirc). Its ability to work with character LoRAs is one of the best I've seen to date, which so often is a problem with so many other XL-based models. test_V1 was pretty good with that as well. Hands look fine most of the time, but still see incorrect digits sometimes (ie, 3 finger and a thumb on one hand and 5 fingers plus thumb on the other).
Anyway, excellent work and thanks again for sharing!
Thanks.
From what I'm being told and finding out, your always going to have the issue with hands. I still feel there is a way to correct that. I am one that I go for perfection all the time so Im never happy with my end results but it still fun. I never use neg. prompts, I just don't believe in them.
I'm wanting to learn to train models and not merge them, I think that gives you the best chance for good hands. I think if everyone keeps merging models then you are just passing around hidden issues. Everyone knows way more about all this than I do, I just hope I can bring a new mindset into all this.
Thank you for testing my models and I really do appreciate the feed back
@jrrytng467 Think you're right about training from scratch. Just guessing, but when your're merging different realistic models that have already had various forms of tuning for hand correction baked-in, then maybe the info can clash instead of improving things further. Sorta like how I always use a detail tweaker LoRA at low strength, but then I'll come across a model that has a similar LoRA merged in strongly and then the effect is compounded to the detriment of the image.
Plus, you have a really good eye for something capable of high realism and keeping a proper but interesting level of depth and lighting without being unnecessarily styled, all while keeping a standard, easily usable setup (no super-low CFG needed, no 1-8 turbo steps or DMD2 is needed, short prompts work fine, long prompts work amazingly well, a million steps still looks fine, prompt and LoRA adherence is top-notch and DMD2 still steps things up if so desired). Think you could cook up something really good if you learned to train your own model. And feel free to post up early versions. Haven't gotten around to trying P4P v3.0 yet, but I still think the v2.0 you removed had some amazing samples posted iirc.
Just curious, but do you have any PMP v8 by @iamddtla merged in these? Results and behavior seem similar at times. Had some troubles getting good results with many earlier PMP releases, but v8 is an oddly capable and similarly competent realistic XL-based model, even for regular fully-clothed SFW stuff.
@AFD_0 Some of my earlier models might. I have played around with that model before. I like Asian woman and sometime I merge models in just for a different character look. A lot of characters on different models all have a certain similarity to them, I'm trying to get away from that
@jrrytng467 For me, it's more how PMP v8, test_V1 and Hands_Experiment all adhere unbelievably well with any character LoRA I throw at them (P4P v3.0 to a much lesser extent). So many other models lean so strongly into their own faces and characters that they struggle to replicate the likeness from a LoRA. Even some of the most popular models being made often have difficulty with that. But yes, having character diversity and avoiding same-face syndrome on its own is equally important.
With DRAM/NAND/VRAM prices being crazy, it might be a while until I can finally build a new rig that can train a full checkpoint. So for now, I guess training LoRAs and doing merges is a lot quicker and less-demanding, though I'm hoping someday I'll be able to learn how to train a full checkpoint as well.
Anyway, best of luck with whatever you try next!