Suzanne's XL Mix

It's the extra noise and all that enhances the detail in this model. A few samplers can't work with it. If you like Euler, you can usually use Euler a instead of a dpm.

@Picky_Pickles @swedishViking any chance ya'll might want to test out an experimental version of this mix? I'm trying a new method and have gotten some really good results, but I need others with a variety of platforms and prompts to run it through the ringers.

@Picky_Pickles I’ll post up the model soon for testing. So far it looks like it’s just as good as the version 5 models, but better subject renders with many fewer watermarks.

@Picky_Pickles Its up now at https://civitai.com/models/272194?modelVersionId=306802

Try out the test model at https://civitai.com/models/272194/suzannes-xl-testing-mix, and let me know your critique. I thought 5.1 would be one of the best versions of this model I could make, but the new one may be better from some standpoints. There may be a few more things I want to try, but that's the first one listed as 6.0 Test on that model page.

@Suzanne other than running in Python script, I've used only three programs and Fooc0us isn't one of them. I am intrigued by their user experiment promises, but I spend most of my time hacking apart models instead of generating images, so that's not important for me. I typically use A1111 or ComfyUI.

That said, I've noticed the biggest difference in generations using the same model is how the seed is generated or how the model handles prompt emphasis and chunking. For example, the emphasis on (girl:1.3) is calculated differently in Automatic and ComfyUI. In addition, long prompts are chunked in A1111, so you can use more than 75 tokens, but the 76th token is treated as important as the 1st token (and so on for multiples). That can lead to some variety in generations.

But back to the topic -- what's better in 6.0 Test? Is anything worse? I can tweak quite a lot still with the new formula, but detailed feedback will help make sure I get it right when I release it.

@Suzanne Great! I’m in no hurry, I have other experiments to conduct

@Suzanne, what settings do you use with those loras -- strength wise? Do you almost always use it at the same strength for those two or do you vary it a lot when applying them?

@swedishViking It is funny how the smallest tweaks make the biggest differences.

@Suzanne, I'm experimenting with including the enhance-xl-tool in the model. So far, the differences are either so minor that I can't see them or so drastic that it totally changes the composition. I've stumbled on a new way of merging the models that I've used with this mix with a great deal of success so that I may post a totally new model page to open up some beta testing for people to generate with it and see if they think version 5 is better than the new one. But if I do, that new model will have the enhanced tool included as long as the license doesn't forbid it.

@Suzanne @swedishViking I'm trying something different with this model... not sure if its going to be better, but I posted it up as a test model at https://civitai.com/models/272194?modelVersionId=306802 -- try it out and see what you think.

Just giving the people what they want.

It’s very careful analysis and trial and error modifying one block weight at a time at each step. Version 5.1 has extra tweaks to increase the detail and contrast in images, but not all samplers can deal with the extra noise that’s added to produce those effects. Most samplers will work, but others will give you images that essentially look like they were interrupted.

I'm running another test on a method to improve things, but not sure if its significantly better to be a new version. Its up at https://civitai.com/models/272194?modelVersionId=306802 as a test model.

Ah, these are very good comparisons and data. So, the V6 Test method is good enough to become the next version of the mix, as it's an advancement over version 5. Unfortunately, I think RealCartoon XL is always going to be better with the SFW generation unless it's fine-tuned for NSFW images... and even then, it may affect the quality. But then again, that's the whole point of the merged model -- to get as close as possible in quality and aesthetics while adding in the NSFW capability. I'll see if anyone else will use and add information about v6 test, then start playing with it for a new release as Version 6.