https://civitai.com/models/282341 Bros, we have nai3 at home. For real. Full scale finetune of Pony Diffusion 6 with dataset of 1.2M anime pictures: ・Vast knowledge of general, cultural, sfw/nsfw concepts missing in original pony ・Great base aesthetic ・Good stability without sacrificing variability ・3500+ artists styles, few general styles ・Thousands of characters ・No annoying watermarks ・Unique angles, foreshortenings, fullbody-wideshots or extreeme closeups without any issues, pretty backgrounds as an added bonus ・Unmatched performance with tails concepts for your fox/cat/dog/... waifus/husbendos Unlike the majority of PD-derivatives which are just reskin (often just a merged style/tweaker lora), this checkpoint adds a lot of new stuff and changes the game. In 0.4.5 500k pictures from dataset have been captioned with Claude 3 OpusVision, then captions were pruned and converted to hybrid with booru tags. In combinations with accurate approach to TE training, this significantly improves prompt control and allows some funny/useful things like prompting multiple characters without extensions (much stable then before), referring to their traits or outfits, using natural text for more accurate description and so on. v0.4.5 changelog: * Large work on improving stability and anatomy. Works better, less flaws in complex poses, better fingers * New data, characters, artists, concepts * Pyramid Noise has been applyed at final pass of training. Now brightness can be controlled by promt, also it fixes closeups/wide shots and maintains compatibility with existing LoRas unlike standard noise offset. * Abandoning the original quality classification tags and switching to own * General fixes and improvements (Dataset cutoff: 20th May, all requests after this date will be in next version) Features and prompting: Basic: Same as for all SDXL, ~1 megapixel for txt2img, any AR with resolution multiple of 64 (1024x1024, 1152x, 1216x832,...). Euler_a and CFG 5..7 (or drop down to 3..5 depending on your taste), highresfix: anyGAN/DAT, x1.5-1.6, denoise 0.5, upscale works best with single tile resolution no more then 3mpx. Highres fix and further upscale will significantly improve quality, details, eyes, hands, feet, etc. Clip Skip 1 is recommended unless you are using loras that have problems with it, 2 will also work. Quality classification: First of all there are only 4 quality tags: masterpiece, best quality, low quality, worst quality Use 1-2 for positive and 3-4 for negative. Avoid using score_x, source_x, ... etc like in original pony. In most cases they just make things worse, add noise and mess, brake bodies, fingers, change styles and bring back urine yellow-green filter. Originally that was definitely not the best implementation of quality tagging including some training flaws and requiring tons of tokens. On early stages in previous versions it became clear that it's better to introduce new tags instead of using original. At this point after quite a long training they only bring old triggers without serious improvements. Artist styles: Grids with examples in rentry Used with "by ", multiple gives very interesting results, can be controlled with prompt weights. by ARTISTNAME1, [by ARTISTNAME2, (by ARTISTNAME3:0.8),...] or/and [by ARTISTNAME1|by ARTISTNAME2|by ARTISTNAME3|...] Works best in the very beginning of prompt. Can be used ad a wildcard (beware, there is a flaw in sd-dynamic-prompts extension that sometimes wrecks up results when used with batch size more then 1). For majority highresfix/upscale improves quality a lot. General styles: 2.5d, bold line, smooth shading, flat colors, minimalistic, cgi, digital painting, ink style, oil style, pastel style can be used in combinations (with artists too), with weights, both in positive and negative prompts. Characters: Use full name tag same like on boorus and proper formatting, like "karin_(blue_archive)" -> "karin \(blue_archive\)", use skin tags for better reproducing, like "karin \(bunny \(blue_archive\)". This extension might be very usefull. Most characters are known by the name, but it will be better if you prompt their main features, like: karin \(blue_archive\), karin \(bunny \(blue_archive\), dark-skinned female, purple halo, ponytail, yellow eyes, playboy bunny, fishnet pantyhose, gloves Natural text: Use it in combination with booru tags, works great. Use only natural text after typing styles and quality tags. Use just booru tags and forget about it, it's all up to you. And yes, it's still based on pony, so it will be worse in IRL concepts, references or some complex expressions comparing to other checkpoints based on vanila SDXL. Basic negative: (worst quality, low quality:1.1), bad anatomy, error, bad hands, watermark, ugly, distorted, monster correct according to your preferences. Please, do not put tags like grayscale, monochrome in negative, no need to fix washed colors or "yellow filter" here. It will only lead to burned images. 3d in negatives is also not the best choise in most cases. To improve backgrounds, add to negative simple background, blurry background, abstract background but do not forget to remove it if you are prompting something with simple. If you getting unwanted text, effects, bubbles, etc, add to negative manga sfx Lots of Kemonomimi-related concepts: tail censor, holding own tail, hugging own tail, holding another's tail, tail grab, tail raised, tail down, ears down, hand on own ear, tail around own leg, tail around penis, tail through clothes, tail under clothes, lifted by tail (booru meaning, not e621) and many others with natural text. Some reproduces perfectly, some requires rolling. In v0.4.5 works better, less unwanted mindfucks like knots, horse-dicks and so on (unless you prompt it ofc). Dragon/demon/succubus traits like horns and tails also should work better now, but in current version less attention to them, to be done in next. Known issues: ・Some artists have brightness/contrast bias, fixed with prompt or extensions ・Can be better in many ways, some concepts are unstable ・To be discovered, still WIP Not checkpoint issue, but needs to be pointed: Using (weights:1.2) in prompt for all SDXL models requires Emphasis: No norm settings, otherwise there can be broken images with blobs or artifacts. AFAIK Forge does not have this setting by default and some versions even don't have prompt weights control at all. Requests for artists/characters in future models are open. Follow for a new versions. Thanks: Artists wish to remain anonymous for sharing private works; Soviet Cat - GPU sponsoring; Sv1. - llm access, captioning, code; K. - training code; Bakariso, NeuroSenko, T.,[] - datasets, testing, advices; other fellow brothers that helped, and everyone who made feedback and requests. AI is my hobby, I'm wasting money on it and not begging for donations. If you want to support - share my models, leave feedback, make a cute picture with fox/cat-girl. And of course, support original artists. But... 20$ is 20$: BTC: bc1qwv83ggq8rvv07uk6dv4njs0j3yygj3aax4wg6c ETH: 0x04C8a749F49aE8a56CB84cF0C99CD9E92eDB17db if you can offer gpu-time (a100+) - PM. Please leave link/name if using in merges, the model is easily recognizable. Also check out viral licence of original Pony Diffusion XL V6.