Hello, very new to image generation. Downloaded stable diffusion web ui, furrytoonmix checkpoint. Got some promising results, though I am unable to figure out how to clean up the coloring and lineart. It looks very "early AI"-like, almost mosaic. Below I've included litterbox links to the results so you can see what I mean. any help is appreciated

https://litter.catbox.moe/u4xyuv4uj2ryesxg.png
https://litter.catbox.moe/mwp7izfcs1opifov.png

Glad to hear you got everything up and working! Congratulations on your first gens!

There are a lot of different techniques for getting the best quality out of AI gens. You can find a lot of videos and guides online, and there's tons of different advice for all kinds of aspects. Looking at one of your prompts:

clean lines, clean lineart, clean coloring, refine coloring, refine lineart, good anatomy, plump dog pussy, big vulva, canine pussy Negative prompt: sapient, anthropomorphic, on two legs, deformed, extra limbs, incomprehensible, safe for work, anthro, anthropomorphic, unclean lines, stomach bulge, cropped, human pussy, human vulva, happy expression, smiling, muddy colors, blending colors, mosaic Steps: 81, Sampler: Euler a, Schedule type: Automatic, CFG scale: 13, Seed: 4081467157, Size: 504x640, Model hash: 976c5866d5, Model: furrytoonmix_xlV3, Denoising strength: 0.91, Version: f2.0.1v1.10.1-previous-224-g90019688p

Here are some things that came to mind:

1. Each model/checkpoint has pros and cons, but typically newer ones have better quality. Some I'd recommend are the 'Mond models at https://huggingface.co/Xeno443 , which go from anime influenced with StableMond to very realistic with 3WolfMond-LastSDG but are all flexible for different styles. Also all the Nova models are high quality out of the box: https://civarchive.com/users/Crody . Those are less flexible, but there are many different models for different styles.

2. The resolution you use can impact the gen a lot. SDXL models were trained at 1024x1024 resolution, so tend to work best around there (although you can adjust for different aspect ratios). It looks like you're genning at 504x640 . Although smaller resolutions are faster, they can also reduce the quality of the gen because the AI isn't as good at working with smaller spaces. If you go to 2048x2048 or higher that can be too big and the AI can start duplicating body parts, but something around 1024x1024 is the sweet spot (or around there like 768x1344, 832x1216, 896x1152, 1024x1024, 1152x896, 1216x832, 1344x768, etc).

3. CFG (Classifier Free Guidance) controls how creative the AI can be or how closely it will try to conform to the prompt. I've found it largely impacts color saturation. Typical values are from 5-9, with 7 being the default value in A1111/Forge/Reforge. It looks like you're set to 13, which is probably too high.

4. Different samplers like different numbers of steps. Euler a became one of the most popular samplers because it requires a low number of steps, oftentimes around 20 or so is enough. It looks like you're set at 81. That makes the gen take a lot longer, but doesn't do much to improve quality. 20 might be a good starting point.

5. For prompting models trained on booru tags, using the actual booru tags works best since that's what the AI will understand the most. There are extensions you can install that will autocomplete tags for you like https://github.com/DominikDoom/a1111-sd-webui-tagcomplete , but you can just manually check tags on Danbooru or e621 as well. For instance, you have "plump dog pussy" when checking tags something like "canine pussy, big pussy," might work better.

6. Likewise for tagging, the AI will only understand quality tags it's been trained on. Things like "clean lineart" and "refine lineart" aren't likely to improve quality, but they might make you lose color since "lineart" might be tagged on many images without color. For illustriousXL and NoobAI based models the list of quality tags they were trained on is pretty standardized, you can find documentation at https://huggingface.co/Laxhar/noobai-XL-0.5?not-for-all-audiences=true but basically "masterpiece, best quality," in the positives and "worst quality, worst aesthetic," in the negatives should help. You can try the other ones too like "newest" and "highres" in the positives.

7. Typically when prompting, decide on fast settings you're okay waiting for and build your prompt bit by bit so you can see what tags are doing what and what's needed and what's not. For instance you could start with just "dog, feral, female, canine pussy," and see how that looks, then add or remove one thing at a time. After you decide you have a good final prompt, you can increase the resolution and steps and such for a final quality gen.

Hope that helps! Feel free to ask if you have any specific questions. There's also a #gen-help channel on the Discord if you want to ask there. Good luck with all your gens!

Adding to that, you probably want to have denoising strength at 1, unless you're doing inpainting. Civitai also hosts a variety of checkpoint models, along with LoRAs.

I follow a few guides that helps with prompting:
-https://civitai.red/articles/31050/illustrious-noobai-prompting-master-guide (requires an account to view it, which I recommend creating one, also for viewing and downloading models).
-https://rentry.org/fluffscaler-pdxl (for Pony Diffusion XL models, PDXL is somewhat dated, but prompting is similar with newer models, just drop the score tags and source tags if using Illustrious or NoobAI models)
-https://rentry.org/fluffscaler-noob (for NoobAI models, more up-to-date)
-https://rentry.org/fluffscaler-inpaint (for inpainting)

Thank you! I've watched no tutorials whatsoever and only gone off of what the webui has explained. These tips will definitely come in handy.