Meet the revolution! v0.4.0 is here. Anime finetune of pony the way it should be (perhaps the only):

• Vast knowledge of booru tags, especially missing in original pony

• Great base aesthetic

• Hundreds of characters, just prompt your waifu

• 1300+ artist styles, few general styles (styles walkthrough on rentry)

• No more annoying watermarks

• Anything from 1girl_standing_masterpiece to hardcore stuff, great flexibility and lora compatibility

• Outperforms NAI3 in tail concepts (and in many other ways)

New version of my finetune of Pony Diffusion V6 XL. Just pure and full-scale training without any merges of loras/other checkpoints. Huge dataset (~0.7M) gives unmatched knowledge of anime art concepts, tags, artists, characters, etc. Filtering, classification and some special magic gives great base aesthetic, better prompt control, backgrounds, etc. maintaining anatomy and nsfw capabilities of the original model.

V0.4.0 changelog:

• no more freaking watermarks! Well, still it appears sometimes, but mostly comes from original Pony.

• new styles, characters, concepts

• improvements of existing balancing, making them more stable and controllable

• much more stability in existing concepts (like tailjob lol)

• better anatomy, fingers, faces, details

• about 50k of images have been captioned also with natural text using Claude 3 Opus Vision, even this small amount allows to improve natural text and improve general prompt control

Features and prompting:

General styles:

just put some of these in the beginning or somewhere in your prompt:

2.5d, bold line, cgi, digital painting, flat colors, minimalistic, ink style, oil style, pastel style

can by used in combinations and both in positive and negative prompt.

Quality classification:

has been applied with masterpiece, best quality, low quality, worst quality

Score is based on specially designed classifier system. It works along with base model scores (score_9, score_8_up, etc).

Just use

best quality, masterpiece, score_9, score_8_up,...

and

(worst quality, low quality:1.1),score_4,...

for negative

Lots of tail concepts:

tail censor, holding own tail, hugging own tail, holding another's tail, tail grab, tail raised, tail around own leg, tail around penis, tail through clothes, tail under clothes ...

and many others with natural text. Some works perfectly, some requires rolling, some can produce unwanted/enjoyable mindfucks. In v0.4.0 you will have more stable outputs without unwanted knots, horse-dicks and so on (unless you prompt it, lmao)

Artist styles:

Used with "by ", multiple gives very interesting results, can be controlled with prompt weights.

by ARTISTNAME1, by ARTISTNAME2, by ARTISTNAME3,...

or/and

[by ARTISTNAME1|by ARTISTNAME2|by ARTISTNAME3|...]

or just use wildcards. Works best in the very beginning of prompt.

Tags like source_anime, uncensored, spamming of score_x_up and others can give unwanted bias. General style tags from above can be applied along with artist styles to tweak it.

Basic negative

(worst quality, low quality:1.1), bad anatomy, error, bad hands, watermark, ugly, distorted, monster, score_4

Well feel free to experiment, default from pony also works, just add "worst quality, low quality". Avoid unnecessary tags like grayscale, monochrome, 3d, source_pony etc. unless you know what are you doing. Base saturation and contrast are already fine as is, for more flat anime style better use "flat colors and/or minimalistic" in positive (with desired weight).

Backgrounds can be improved simply by adding simple background, blurry background, abstract background to negative prompt, but beware of overusing it.

Use hires fix to improve details, eyes, hands, feet, etc.

Known issues:

• Well, fingers, toes, etc are better than v0.3 and vanilla pony, but it can be better. Refining stage is training right now.

• There are still rare cultural concepts (like lap pillow invitation) that have to be added explicitly for proper performance, will be in next versions.

• Some rare artists and characters may be barely recognizable, please report if you find any.

Still WIP.

Requests for artists/characters in future models are open. Follow for a new versions.

Plese leave link if using in merges, behavior of the model is easily recognizable. Also check out viral licence of original Pony Diffusion XL V6.

Special thanks to my fellow brothers who share same cultural beliefs and helped to make it real.

AI is my hobby, so I'm wasting money on it and not begging for donations (but 20$ is 20$ if you can offer gpu-time - PM). If you want to support - share my models, add feedback, make a cute picture with fox/cat-girl. And of course, support original artists.