This lora can recreate art style and some of the main cast from re:zero season 4.V2 Retrained this time using Ostris v3 adapter instead of distill guidance, and using a validation dataset to find the best checkpoint and micro batch 8 and compile. Also claude had pointed out 15 or so images had incomplete captions. Trained on musubi fork.This new version is way better; it can do more characters consistently and the lines and colors are more stable than V1. I also see it works better with ref model. I had prepared more data from the remaining episodes in the season but forgot to put them in unfortunately, though I think they're not needed.V1Trained on 462 screen captures from only season 4 (ep1-16) of re:zero . So consider this a lora which can recreate some of the main characters from re:zero season 4. Some pre-season 4 characters are in there but do not have enough data to recreate them well. I will add more characters in the future and as I learn more want to retry this lora. Trained at 1376×768 on Musubi Tuner Fork. This was put together using my new lora dataset preparation tool, and captioned using Gemini 3.5 flash. It took around half a day to prepare everything and I just took my dataset from the KREA version of this lora and converted the captions to H3 syntax.It trained relatively quickly, I actually tried training it 2 times. first with automagicv3 and then with adam8 optimizer and automagic was better overall. Krea managed to get almost every character perfectly, but H3 still I don't know how to get multiple characters right, subaru, emila, rem/ram, and reid (the first few probably because they're in the base model somewhat) can get done pretty much fine but all the other cast not so much though I think it gets their clothes right like beatrice is weird looking but outfit seems fine, and same for julius and anastasia. I considered rank increase but I think its just the method of training the model with how it its distilled makes it hard. Hopefully we can learn more about training h3 to make multiple characters better in 1 lora. I think you can use this as a base and just train the individual characters in a few hours. You can also just use the reference workflow with the FLV2A model with my KREA as reference input or just i2v and it will work better.As far as creating the style of the show, it picked that up greatly. And I think things were fine around 4k steps in, but I like the 15K checkpoint the most, started to see it burn in wrong details for characters after that. Also some things like hair ribbons, or hair styles may be a bit burned in, I think in next version I need to carefully caption all ribons and hair styles. I'll upload the 4k checkpoint for you try as well as an optional file. Most of the examples are t2v 1-2mp res and using 20 steps euler beta 57 with comfykitchen. Not yet tested with any turbo loras.CONSTRUCTIVE feedback always welcome, and I like to see your guy's samples.<h3 id="style-trigger:">Style trigger:</h3> 2d anime style, rezero styleYou can just create in the style of the show, don't have to use the characters if you don't want to.To trigger the characters, use the given character tag and a simple description of their appearance and outfit. For H3 the character tag is quite important and so is the description. For help prompting I've provided for all the characters below:char_subaru: char_subaru has short black hair, brown eyes, he wears a dark hooded jacket with an orange collar and a green shirt underneath.char_emilia: char_emilia, with long silver hair, elf ears, and purple eyes. She wears a purple full body leotard and a white cloak with a circular gold brooch at the collar.(You can also prompt her without the cloak, you may have to play with the prompting a bit to get the body suit leotard fully covering her chest)char_beatrice: char_beatrice, a young short girl with blonde hair styled in twin drills adorned with large pink ribbons, blue eyes, and fair skin. She wears a pink and red frilly dress with white trim, a large pink bow at the front, and a white collar.(her eyes need work, going to improve her in next version)char_ram: , char_ram a young girl who has short pink hair, red eyes, and wears a teal hooded cloak over a white and green outfit with a dark corset and ruffled skirt. (You can also prompt her in french maid outifit)char_rem: char_rem, a young girl with short blue hair covering her right eye and blue eyes, wearing a blue maid uniform with a white frilly collar, white headband with round ornaments, and a pink ribbon accessory in her hairchar_julius: char_julius, a young man with purple hair and amber eyes stands outdoors against a bright blue sky with white clouds. He wears a white knight uniform featuring a long white cape with a hood, red trim, and black cross-straps fastened with golden clasps. His hand, covered in a white glove, grips the golden hilt of a sword at his hip.char_anastasia: char_anastasia, a young girl with light purple hair and blue eyes stands outdoors at night near a stone balustrade. She wears a tall white fur hat adorned with a pink ribbon and star accessory, a white hooded cape with fluffy cuffs, and a white outfit.char_meili: char_meili a young short whol has pale skin, green eyes, and medium-length light purple hair styled with a long braid resting over her left shoulder and a purple flower-shaped hair ornament on the left side. She wears a dark grey dress with a white ruffled hem, puffy white cuffs at the wrists, and a dark brown hooded cape fastened at the neck with a gold clasp and pendant.char_shaula: char_shaula, a tall girl cwith long dark brown hair tied in a high ponytail and a braided section, amber eyes.. She wears a black strapless bikini top with a bow and star accessories, black shorts with an orange belt, a dark cape with orange lining and bat-like details, and thigh-high boots.char_reid: char_reid, a man with long red hair and a single blue eye, wearing a black eyepatch over his left eye. He is shirtless, revealing a muscular torso and defined abdominal muscles, and wears a red robe draped loosely around his waist with dark trousers underneath. He holds a pair of chopsticks.<h2 id="the-following-characters-in-the-dataset-but-results-vary:"> The following characters in the dataset but results vary:</h2> char_rui: char_rui, a young short girl, with blonde hair in a long braid, blue eyes, and sharp canine teeth. She wears a white frilly dress.char_roy: a young boy, char_roy, with long brown hair partially covering his face and wide green eyes with small pupils, wearing a dark top. His mouth is wide open in a manic expression, revealing rows of sharp, triangular teeth, and his extraordinarily long tongue lolls out of his mouth, coated with white saliva.char_lye:a young boy char_lye with long brown hair, dark skin, and a worried or cunning facial expression. The boy has one large green eye visible, while the rest of his face is shadowed by a section of his hair covering the right side, along with dark shading across her forehead and brow. He has a small, sly smile with a single sharp white fang protruding from her upper lip, and a thin, worried eyebrow line above his visible eye. He wears a dark garment with red abstract patterns and a purple bow or accessory at the collar. (I think lye/roy got mixed up in the dataset)<h3 id="the-below-are-also-tagged-but-not-enough-data-to-remake-them-well:">The below are also tagged but not enough data to remake them well:</h3>char_grarfiel, char_frederica, char_reinhard, char_roswaal

Re:Zero Anime Style / Character Lora - H3

Retrained this time using Ostris v3 adapter instead of distill guidance, and using a validation dataset to find the best checkpoint and micro batch 8 and compile. Also claude had pointed out 15 or so images had incomplete captions. Trained on musubi fork.This new version is way better, and can do more characters consistently.

Files with SHA256: ed30172afcf776a779836afc46bd5bf4de10a486b830626910fa9a11e3611d77