The Method - SDXL

After playing with ComfyUI for some months, i'm now using SwarmUI to generate pics.

Obviously, the models used are SDXL Pony models.

I convert what i often use to TensorRT models, which speeds up generations considerably by optimizing models for the specific GPU you have (about 40% gen time shaved off). Unironically, thanks, Nvidia.

I do not use LoRAs, since they tend to fuck up faces comprehensively. If i never see another of the horrendous default SDXL/Pony faces again, it would be too soon.

What i do use is Embeddings (=Textual Inversion). Since i do mostly 21:9 format, i was plagued by doubling of prompt subjects (or quadrupling etc), elongated bodies (how many navels can you fit into one belly?), and/or pictures split into multiple columns (the column_lineup tag gone rogue).

Stable Yogi's positive and negative embeddings were a game-changer for me. Not only do they put a stop to column_lineup nonsense,. they also stop stretched bodies and other crap SDXL tends to produce at 21:9.

Which led me to trying out other embeddings, most notably those of Void91 aka Hikari, who does the HDA (Hentai Digital Art) series of embeddings.

Although they are made for hentai (which is so totally not my thing), nearly all of them work well in realistic pony generations, with only a few producing unwanted effects (which can mostly be counteracted with some choice danbooru tags in the negative prompt.

I use an extensive array of wildcards in my prompt to ensure a varied output (I'm easily bored). I re-wrote my existing wildcards after (reluctantly) studying the danbooru 'sanctioned' tags, which helped more than i like to admit.

And that's about it, none of the time-consuming stuff (inpainting etc) - at my current settings, i can generate a picture like the one to the right every 20 seconds.