[IL] Refia (Final Fantasy III 3D) / レフィア ( ファイナルファンタジーIII 3D)

https://civitai.com/models/1623613/il-refia-final-fantasy-iii-3d-iii-3d

A model for Refia, the best part of the Final Fantasy III 3D remake. A Final Fantasy crossover needs to give her a rivalry with Firion: She hates toads, and Firion is master of the Toad spell. Also shame on Wizards of the Coast for skipping including her when the designs of the later 3D remake of 4 was able to be included under the same mainline games only scheme.

Head: F_F_T_R, orange hair, long hair, flipped hair, earrings, brown eyes Freelancer: R_F_L, white dress, long sleeves, blue vest, black pantyhose, thigh boots, fold-over boots Devout: D_U_S, cat hood, hood up, black neck bow, white hooded capelet, pink dress, white sleeves, thigh boots, fold-over boots (beware views below the waist on this one can randomly get slapped with a PG-13 rating like any skirt/dress that doesn't cover the knees. Note: Illustrious can "Red Mage (Final Fantasy)" and, with a few supporting tags, "White Mage (Final Fantasy)" by default (Black Mage doesn't work though).

Data Count and Commentary: 27 pics in training: 2 (technically 3, see next) pics from the pre-rendered 3D intro, 2 pics from Steam Community items (one of which is just a pic from the 3D intro but at far higher quality), a 3D promotional render (technically one of the Steam Community pics is too), 3 screenshots of the PC version of the remake, 2 official art of her from DS, a pic of her from Pictlogica, a pic of her World of Final Fantasy model, a pic of a promo art for the PSP release that was split into 2 (each side of her reflection), and a bunch of fanart. 18 pics were Freelancer, 7 were Devout, one was a White Mage (shots from the intro), and one was Red Mage.

Refia suffers from having a true mess of training data, even compared to the much more obscure characters I’ve done before, which is likely why nobody else has done her without synthetic data yet: Loads of styles with their own distinct differences in design (her hair color is all over the place, and the 3D intro has straight hair), a large chunk of the data is chibi, she’s frequently seen in alternate costumes, and the best stuff is 3D. A true “dog’s lunch” of a training data set that I’m surprised works as well as it does (clasps (?) on the dress appear to be pretty inconsistent though). The benefit, and negative, of having this varied a dataset however is that she has low "inherent style" so she'll come off kind of generic on her own, but pairs well with any style you stick her in.