The last test I did focused too much on the character and missed the style of the show, so I saved up some buzz and re-made it! This one took a lot of buzz, but is a significant improvement over the previous version. Do note the example images are just some initial tests and I am still experimenting with best prompting practices! I will update with notes and example images if I find ways to improve.The model is now much better at reproducing the style of the show, and has some understanding of some characters and concepts too (although not as broad of an understanding as the pdxl version, so do not expect to reliably get accurate outfits, or even characters other than tarte if you do not describe them well in natural text)Trigger words remain largely unchanged, and self explanatory for the most part: sa+[character name], outfit and hairstyle descriptions, with the most prominent being assigned a unique no-spaces tag, the rest is up to how well you can describe it to flux. Due to current limiations of flux training, there is now fewer tags it knows compared to pdxl, and it is recommended to weave them into a natural language prompt for best results... and do not expect it to work as well as pdxl in terms of style with a complicated scene unless you are a true prompt god.Main tags that were trained are: sastyle, -general style trigger, main use for this loraoptional, weaker trigger words that you can try to get characters:saTarte, -Tarteorangedress, maid, mob cap, servantdress, assassinsuit, twintails, pink bikini, hair rings, long hair, blonde, etc -appearance customization optionssaMaha, mahahairclip, saDia, white hair, brown hair, double bun, saTarteY - young tarte from beginning of show and flashbacks (separate tag since there is a larger difference in her appearance compared to the others)saLugh, white hair, black hair,

World's Finest Assassin Anime style FLUX (as well as a few charcaters and outfits)

Initial release of style version that also tests how much character and outfit understanding I can cram in to one flux lora. Style transfer is reasonably good, but characters may require several tries, especially in complex scenes. It is still not perfect, but it is unlikely that I will be making a v2 any time soon with how expensive flux training is unless this gets popular enough to offset the training cost, which I highly doubt.

Files with SHA256: cf1c1e0141a4697b9dd149bdb22a760d452b19330d983fc6d849464b81a9457a