Suzanne's XL Mix

@Suzanne, it was a quality tag used in SD 1.5 training, which basically means the resolution should be so high it's crazy. When I generate images, I have a "basic" style that I always apply that has that in there. It keeps me from having to type a few common positive and negative words over and over.

@shapeshifter83 Yeah. There's lots of issues with some of the prompts I'm using. I do checkpoint testing by using the civitai prompt wildcards, and though I majorly cleaned up the file, many are still misspelled or ineffective. I set it to infinitely generate on the wildcards and then come back and pick out ones I like for the showcase. But I should really go and refine the ones I'm using and pay more attention.

@Suzanne if can I determine which unet block the concept is in, I can merge more weight from that particular one.

@Suzanne working on a new version of the mix that seems to be able to accomplish the lighting like this! Testing is positive.

@Suzanne I'm looking at ratios right now. Take a look at the grid and see what alpha and beta values (the numbers going up and down) hit the sweet spot for aesthetics. Give me a range for both where the images are good for the look. I'll have a few more grids as they finish running. https://civitai.com/posts/1163257

@Suzanne model mix has Better than words merged. I'm going to redo one of the base parts and try again. Lora's are odd to mix as, generally, they only appear when summoned (using trigger words) and can significantly change the style when that happens. It can make a realistic model suddenly go anime and vice versa. But I'll take a look at those you mentioned as anything without a trigger word works well as a poor man's fine-tuning when mixed in.

@shapeshifter83 I'll look into that. I found that the All-in-One and other misbehaving loras work if you merge it to the first model and give it more weight than you think it'll need. Then, use that merged model in the mix as a base. If you use a mbw after that, you can give more weight to the blocks that contain those nsfw concepts (which seem to be those close to the middle). Or extract the weights of the Lora after that merge and use it as a new Lora to merge instead of the original one.

@Suzanne I think I've got something. Can you post three links to some of your images on the site with full-generation data? I'll use them to make more grids with the new custom alpha and beta values and let you pinpoint the aesthetic using prompts, settings, and seeds you've used before.

@shapeshifter83 It's ready now, try it out and let me know what you think so I can consider what to change in the next version.

I tested with these and many others. Looks like I narrowed version 5 down to two models. Both give very good results and fall in the range of tests you picked in the grid. In my opinion, they both seem to create very good-quality images, listen carefully to prompts, and aren't as overly horny as the other versions based on the protogenesis model -- but they still will create every bit as good nsfw images. I'll narrow it down to one model and post it as version 5 later today.

Version 5 is here! Try it out. Let me know what your feedback is to improve the next one.

@Suzanne, Have you had a chance to use the newest version yet? Any feedback?

@Suzanne I think we can make those improvements. Composition (how the components and subjects are positioned and arranged, the POV, the focus, the lighting, etc) is hard to change. It's largely based on the training data the models were trained or fine-tuned using. Aesthetics (the look and feel, art style, sharpness/softness, contrasts, color, etc) is a bit easier because once we get the model close, we can just make changes to things like level of detail, the contrast, lightness/darkness, and color tone by tweaking the input and output points of U-Net. It's really fun to look at how it makes differences in images from the same seed when you make the model increase the level of intricate detail or the lighting contrast. Adding additional noise makes the model come up with extra components from the noise as it diffuses the chaos.

So, the simple answer is yes. We can easily make complexities for the model to sharpen into additional details and we can increase the contrast and to create better definition between the light and dark. It was actually the step I was planning on trying once I was sure the bulk of the model was producing the results we wanted.

Any more tips on prompting and settings? I can write them up to post on the model page.

@swedishViking, that is to you as well -- any usage tips? I use Automatic1111 and ComfyUI, so I am afraid I don't know much about the best settings for your UI.

@Suzanne Here's a grid of adjustments to the model. Look at the numbers along the left and tell me which series of numbers is consistent for the adjustments you want to see in the images. https://civitai.com/images/5475487

@Suzanne New grids are posted at https://civitai.com/posts/1198107. You can make comments directly on the images to note things. There are 8 variations using a variety of prompts. The first row is the original model. The second is merged with an experimental version, there should not be much change in composition, but the details and contrast may show more differences. The last is a straight 50/50 merge with another model. The first column is the current level of adjustment (so the top left-most image should be exactly what is generated now by version 5). Use this image for comparisons. The second row is a typical adjustment I make if the composite models used in the merge do not have much complexity and detail in the images they create. The sampler used is DPM++ SDE, which should be identical to your dpmpp_sde_gpu. Anyone else looking at this comment is welcome to evaluation and comment as well!