• Thanks for the original creator (welcome back! おかえりなさい)

• Main target: generate accessories accurately without using ControlNet

• Focus on fidelity of the outfit rather than others

• bodysuit in navy blue, black color (and more?)

• Optional chest emblem

• Can generate characters with no mask

• All images in dataset are AI-generated and edited by me (see below)

Usage Restrictions

• EN: Non-commercial use only / 日: 非商用利用のみ / 繁: 僅限非商業用途 / 簡: 仅限非商业用途

• EN: No requests / JP: リクエストは受け付けません / 繁: 不接受任何請求 / 簡: 不接受任何请求

Usage

• Trigger word: zofa soldier

• Recommend weight: 0.9-1

• With white mask (full body):

zofa soldier, [navy blue/black] bodysuit, lxmask, [navy blue/black] lxmask, foreheadjewel, chestjewel, black gloves, black belt, buckle, wmask, white wmask, closed mouth, knee boots, high heel boots, black footwear

• Without white mask (full body):

zofa soldier, [navy blue/black] bodysuit, lxmask, [navy blue/black] lxmask, foreheadjewel, chestjewel, black gloves, black belt, buckle, red eyes, empty eyes, looking at viewer, colored skin, [blue skin/grey skin], visible expression, knee boots, high heel boots, black footwear

• delete tags on template that you don't need

• (case-sensitive) add chestjewel to turn on chest emblem; add chestjewelOff to turn off if needed

• (case-sensitive) add wmaskOff to turn off the white mask if needed

• sometimes you can use (chestjewel, chest crest) together

• use Hires.fix for better results

• use this assist LoRA made by me (v1b) to fix the neck problem of unmasked characters

• In some cases (e.g. [con]fusion of forehead emblem and white mask), use with this Zofa Soldier LoRA (in low weight) may get better results

• see my images for more examples

Tag Explanations

• [color] lxmask = the enclosed full head mask

• foreheadjewel = zofa emblem on forehead

• chestjewel = zofa emblem on chest (optional)

• white wmask = the white mask on the enclosed mask (removable)

• remove "buckle" when generating image from behind

Production Details

• Dataset: Edited AI images (8 full body, 1 cowboy shot) are used, and then cropped into 44 images, all in 1536x1536 resolution

• You can download and use it on your own project

• Process: AI images generation> Preparation of images for Canny preprocessing using SAI > Images generation with ControlNet (Canny) > Manual background removing, collage, color adjustment using Photoshop > Upscaling to remove artifacts > Cropping

• Used LoRA: Models on CivitAI including this one, and a flawed private Zofa Soldier LoRA

• Extra Assist: I used NanoBanana to generate reference image from original ones (e.g. outpaint the rest of body part, generate view from behind and side...). I also asked ChatGPT for advices on tagging

• AI Tools Used: ReForge, ControlNet (Canny), NanoBanana (Gemini), ChatGPT

• Painting Tool/Image Editing Tools Used: SAI, Photoshop

• Trained on CivitAI