I was on Flux and noticed there are many other choices in AI models. Most create low-resolution images of 512 x 512 pixels.
I did a comparison of the ones that are higher res, all of these are at 1024 x 1024 which may be upscaled with Krea.
The reason for the comparison is the poor wetlook results I found in Flux. For my comparison, I used the same prompt for each image and did 4 with each engine to get a good average image from each model. The prompt included the young woman being wet and also with foam on her as well. Each model seemed to handle things in its own way.
Here are my observations regarding generating wetlook. (including how wet other substances may look, such as paint, syrups, slime, etc.)
Flux: Wetlook is very poor, both on skin and fabric. JuggernautXL X: Wetlook is slightly better than Flux, but not great. RealVis XL Ver 1: Wetlook is much better, but other details are sometimes wrong. RealVis XL Ver 2: Similar to Ver 1 RealVis XL Ver 3: Similar to Ver 1 RealVis XL Ver 4: Wetlook is vastly improved over previous versions.
For me, the winner is RealVis XL Ver 4 for anything involving wetness, whether clothing, skin, and whether water or a wet substance.
While comparing, I also wanted to include a similar prompt, and how different Ai software handled it.
You may recall a series with Francesca on a bed that looked like a giant pancake, and with syrup poured on.
The images were made with Bing and upscaled with Krea, but I could never seem to get her dress looking syrupy, just the cover beneath her.
Then, using Kling, I could get better detail, sticky hair, face, dress, etc. but only from the boobs up. I can not seem to get a full-length shot of her on the bed.
Finally, with Flux, I have better control, with her hair, skin and dress looking saturated in the sticky syrup. I can even have her partially (or completely) remove the dress.
So I'm using less and less of Bing, with all its restrictions, Kling mostly for video work, and Flex for most images because it's fast and uncensored.
Presumably you are using the base flux model Vs Juggernaut XL and RealVis XLwhich are finetune SDXL models which have been enhanced with additional curated images.
The prompting requirements for Flux and and SDXL are different. Flux works significantly better with longer descriptive natural language whereas SXDL still prefers key words.
You could likely significantly improve the flux output with better prompting and even more so using Loras which are smaller finetunes on concepts, objects etc. Longer term, I would expect full finetune models of Flux to blow SDXL out of the water so to speak. The creator of Juggernaut is working on a flux model at the minute.
I don't doubt it as I'm no expert. I just noticed a few tendencies and thought I'd share them. I'm glad you're here to shed light and share news on developments. (such as the Juggernaut people working on a Flex version)
I like the image detail in Kling, but for the life of me, I can't get the camera to back away no matter what terms I use, such as 'long shot', 'wide-angle', 'full-length', etc. I've tried the Bing tricks, such as mentioning shoes or anything that would back the camera out, but all that seems to get ignored.