"DAMN!" (c) - my first thought when the finetuning ended and I saw the final results. This model is a finetune based on PonyXL, that aims to bring it to the level of realism of models like Juggernaut and etc. I used a dataset of 500 pics that was carefully curated, captioned and edited by me, making sure to leave no room for mistakes. Still in the process of baking the model, so more versions to come.
Make sure to read the description, there's a lot of important tips.
What makes this model special: ◦ Incredible skin texture. I've made sure to train it into the model as best as I could at this stage of finetuning. ◦ All the original concepts are preserved, lots of them are enhanced to suit the realistic look better. ◦ Suitable for both NSFW and SFW stuff. Use appropriate tags in your prompts. ◦ The amount of watermarks is drastically lowered compared to the base ponyxl model.
◦ Support from the creator (me, duh). If you want me to add some concept into the model, make sure to write about it into the discussion, and I'll see what I can do. TO GET THE BEST RESULTS: 1) You may or may not want to use the "score_9, score_8_up, score_7_up, score_6_up" prompt. The decision heavily depends on your tastes and the rest of the prompt, so I recomend trying with and without it. (Tip: I feel like without "score_9..." prompt pics come out looking more candid/*******, while with it they look kinda staged, but not in the bad way. Just a different vibe to them). 2) Make sure to not use the "quality" tags like "worst quality, lowres, bad anatomy, etc." in your negative prompt. That's right, with this model there's no need to do that, moreover it will only ruin the quality of your gens. Only add something into the negative if a thing you don't want to appear in the picture keeps doing so every time ("depth of field, blurry background" for example). 3) Depending on your composition, you may want to use adetailer/highres fix.
RECOMENDED PARAMETERS:
◦ Sampler: DPM++ 2M Karras / DPM++ 2M SDE Karras / whatever is your favourite, just do not use Euler A.
◦ 15-20 Steps.
◦ CFG: 5-8 (lower for a more washed out/vintage look, higher for more clarity/contrast). I personally use 6 almost all the time. ◦ The model was on trained on 1:1 (square), 2:3, 3:2, 3:4 and 4:3 aspect ratios, so it would help the quality if you used these ratios (although that's not necessary - it behaves nicely on other ratios as well). ◦ Highres fix: Upscale by 1.4-1.6, denoising strength ~0.4, upscaler - 4x NMKD Superscale or 1x ITF SkinDiffDetail Lite v1.
KNOWN PROBLES AND SOLUTIONS: 1) Bad quality of faces on distant shots - use ADetailer or inpaint faces manually. Don't forget that it's a model that was trained on top of a heavily anime/cartoon-biased model, meaning that even more training is required. Will be fixed in the next versions.
2) Bad hands - use ADetailer or inpaint. This model is not immune to the common SD problems just yet, so the hands can be bad in some gens. Will try to fix it, but I'm not sure how to do that yet.
3) Same face - two things can be done: either remove/lower the weight of the "score_9, score_8_up..." prompt, or try specifing the appearance: describe nationality, age, eye color, etc. Already working on improving the variety of faces.
The model was trained on a merge of Pony Diffusion V6 XL (85%) + Everclear PNY (15%) by Zovya, so it might have a little bit of common DNA with the latter. Finetuned with OneTrainer.
