It follows prompts well, it will create images of beautiful women, has 4 fingers and a thumb on both hands. Tell me what this model does?
Description
This is a fix of version V2, I got everything back I wanted with very crisp detailed images. I'm sure you will like this one
FAQ
Comments (20)
I will say this model might be a good model to play around with or even use merging with another models but I feel it lacks a lot of things my test model does. Ive lost the Anime touch, and the 4K crispness. Images seem a little out of focus or cloudy. I think Ill leave it up for awhile to let people play around with, and take it down at a later time. Please leave a comment or feed back please
Looks good. Please leave it up for a few days so I can try it out!
I will, I fixed it though and going to post the new version with it. Im very please with the new results and Im sure you will as well
@jrrytng467 Really wish you would've left v2.0 up for a bit like you said you would. Found it backed up somewhere else, so guess I'll try that. v3.0 looks good as well, but I remember the v2.0 samples you posted looking better for some reason.
@AFD_0 I apologize for that. I left it up over the weekend, I guess I should have asked you, i did say I would leave it up for you.
@jrrytng467 No worries! Hopefully I'll be able to try out v2.0 in a day or two, but v3.0 is a really good model too. Can't decide if I like P4P v3.0 over test_V1, but they're both really good! Think test_V1 has a little better realism, but but P4P has a very nice 3D depth to it without being CGI or cartoonish. Prompt adherence maybe seems a bit better on P4P, but think test_V1 handled LoRAs much better. Results seem like a recent/modern XL with a bit of slightly older Pony mixed in. Very similar output to test_V1, but definitely works a bit differently. Contrast and lighting seem improved quite a bit, but can still be not-as-expected when working with various character LoRAs (pulls less subject likeness vs test_V1 and bumping the strength or emphasis pulls too much from the color, contrast, image quality, lighting, etc of the LoRA like it's a style). Background detail might've been a bit better in test_V1 as well. P4P v3.0 seems a bit simpler in that regard. Just my observations and both are incredibly good merges. Both keepers and think I might even try a simple A/B merge and squish them both together, maybe ~65% P4P v3.0 with 35% test_V1.
Anyway, very excellent work on both and thanks again for sharing!
@AFD_0 OK, but just beware, my models started with a base model Reality BoundXL faceSmash v8.0 reupload (creator is ironically_irascible) If you read his "about this version" doing a Clip merge was what I found out was causing my issues. The model by its self I thought was great but the further I got into merging the more issues I started having which was coming from the base model. For me, this not about pride or anything, I'm learning and having fun and if I can give back in any way, to me, that's whats it about. So with that said, thank you for the complement and you're welcome to enjoy anything I have to offer.
I think version 3 is a much better model but you decide. This is a big learning curve for me just uploading files is a little tricky for me
Does it have some Illustrious or Pony inside? I'm getting a bit of a "sameface", and a bit exaggerated proportions (heads and eyes that are slightly too big). In addition, images come out oversaturated/overbaked, need using extra-low CFG.
Ive been working with only trained SDXL models. This one I merged 3 at the same time. My test_v1 was my base, realismEngine_v10 in second position, and realityBoundXL_facesmashV80Reupload on the end giving an equal mix of the 3. The settings I use on all my images are always the same. CFG is usually always 5 or 6. I was batching images 5 at a time and would see little things not right but GPT told me to stop doing 5 at a time, that was why. I do 3-4 at a time now that helped. The only thing I could suggest is look at the setting in my images I use, and copy and paste someones prompt and see what you get. If that helps then you can work you way out from there. I'm just learning so I cant really help to much but I do really appreciate your feedback. Please continue to do so
I also always set sample method to DPM++2M and Schedule type to Karras, I think Sample method will always play a big part. Some models I have to use LCM to get it to work and the CFG has to be turned way down to 1, which I don't like to do so that's why I always stick with what I use. Its simple and has room to play
So I just used a prompt you had on one of you images on your profile. I seen issues with mine but if I used the same seed I seen better results. I will post 2 images without the seed and 2 images with the seed. For me Ive had issues with just using a prompt which is weird when it looks good to me. Most of my images come from other peoples prompts from there image I like.
That explains it. The author of realityBoundXL_facesmashV80Reupload messed up its CLIP in v7 and v8, and pretty much ruined realism in that checkpoint. Your test_v1 was good and realistic, but the addition of realityBoundXL_facesmashV80Reupload inherited its problems and messed up your CLIP - hence the sameface issue and anime-like disproportions. Here's my comment on that checkpoint at the corresponding page: https://civitai.com/models/1815169?modelVersionId=2461971
"With all three editions of the V8, I am getting the infamous "sameface" issue. It really likes generating just one female face; there is simply no variety. Is there some Poiny or Illustrious in it?"
My recommendation: if you are after realityBoundXL, pick v4, v5, or v6 (these are the most realistic). v7 is ruined, v8 is ruined some more, and v9 began returning to normal.
@civit77899 Thank you for that, that does make scene. That model is awesome at following prompts which is what drew me to it. I will try some of his older version. I do feel like I'm stuck in a loop right now
@jrrytng467 That's the thing. You use Pony/Illustrous CLIP, and you get better prompt adherence - and the Pony-face/Illustrious-face and anime-like body proportions as an unwanted bonus. There's no way out of this loop, it's either/or, proven multiple times, the latest attempt being Snakebite and Snakelite checkpoints. The author of BigLove started by merging realistic Pony checkpoints with SDXL, but all of his latest checkpoints are trained finetunes.
@civit77899 Well, I am learning a lot from this feedback so thank you very much. Back to the drawing board but I need to rethink my approach a little
I just posted a new comment on here and I would like you to read it and tell me what you think. I would like to thank you for your help with my issue, really. I would not find things out like this with out the help of people like you
@jrrytng467 I redownloaded the model and gave it another run - just in case anything changed. I still think it's a lot of Pony in it. Your test V1 is still the strongest model (honestly, it beats many existing merges on this site already). Not this one though (personal preference of course). I am also testing your other model now (namegohere), will post on the corresponding page.
Please take a moment after you've tried this model V3 and tell me what you think over all. I'm seeing clear skin images coming back, I'm happy to see that.
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