I trained a locon for these nine Bocchi The Rock characters I could gather a sufficient dataset on.
I did prune the appearance tags making the girl's name their respective token, but sometimes a slight push helps in accuracy, so here are some overly verbose examples, the green ones are an example outfit:
gotou hitori, cube hair ornament, pink hair, straight hair, blue eyes, pink jacket, track jacket, black skirt
gotou futari, pink hair, blunt bangs, short hair, serafuku - no previews
gotou michiyo, pink hair, single bun, blue eyes, apron ! Sadly the dataset for Michiyo was quite small and mostly consisting of one artist, so you might have to drop the locon strength to 0.7ish for her.
kita ikuyo, green eyes, red hair, one side up, shuuka high school uniform, cardigan
yamada ryou, yellow eyes, hairclip, blue hair, shimokitazawa high school uniform, black sweater, collared shirt
ijichi nijika, ahoge, red eyes, blonde hair, sidelocks, side ponytail, shimokitazawa high school uniform, green coat, white shirt
ijichi seika, ahoge, red eyes, very long hair, straight hair, striped pants, choker, black shirt
pa-san, green eyes, blunt bangs, long hair, colored inner hair, black hair, purple hair, earrings, sleeves past fingers, dress, gothic
hiroi kikuri, pink hair, purple hair, hair over shoulder, single braid, purple eyes, sharp teeth, green dress, cleavage, sleeveless
Description
FAQ
Comments (7)
what model is Counterfeit-V3.0-real-fp16?
where can i download it
It's just https://civitai.com/models/4468?modelVersionId=57618, Counterfeit v3.0 but pruned to actual fp16 without junk (so the model's 2gb not 4gb)
To elaborate a bit, only two valid sizes for SD1.5 checkpoints should be 2ish gb and 4gb, anything more than that contains data that is useless for inference (and LoRA training too, probably), and the size can be pruned down using a script or the model toolkit extension. The outputs should stay the same.
@Aotsuyu oh, understand
Could you please explain how the prompts ADDCOM and ADDCOL work? I've been using the same prompts for my LORA, but the character attributes seem to get mixed together. Are these prompts exclusive to LORA, or can they be used with other models as well?
How can I prevent the features of two individuals from getting mixed together?
Oh yeah, I should have probably specified that those images were made using regional prompter (https://github.com/hako-mikan/sd-webui-regional-prompter), getting 2+ distinct characters is "possible" without extensions but it usually fails/requires inpainting.
Using regional prompter you can prompt differently for each part of an image, and ADDCOM, ADDCOL, ADDROW is syntax specific to it.
The extensions should work on any model/LoRA.
@Aotsuyu I understand now. Thank you very much for your response.
Details
Files
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.

















