TW3K Portrait LoRA (`tk3k_portrait`)

An SDXL LoRA that generates Total War: THREE KINGDOMS–style character portraits in the look of the SAD modpack art. Use it to create new in-style hero/officer portraits on a clean near-white background, ready for upscaling, background removal, or export into the mod's character art structure.

Base model: stabilityai/stable-diffusion-xl-base-1.0 (required)
Type: LoRA adapter for SDXL
Trigger word: tk3k_portrait
App / full pipeline (web UI + export tools): https://github.com/Ironictw2st/tkart-portrait-generator

⚠️ Non-commercial fan content. This LoRA was trained on art from the Total War: THREE KINGDOMS SAD modpack. The game and its assets belong to their respective owners. Released under CC BY-NC 4.0 — personal / non-commercial use only. See License.

Files

File	Size	Notes
`tk3k_portrait_v3.safetensors`	~436 MB	Recommended — final epoch (v3). This is the default.

Download the file and (if you're using the companion app) place it at output/v3/tk3k_portrait_v3.safetensors.

Usage

Easiest: the web app

Use the TW3K Portrait Generator — a local Gradio UI with generation, upscaling, background removal, and mod-export built in. Follow its README to set up.

Directly with 🧨 diffusers

import torch
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16,
).to("cuda")
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

# Load the LoRA from this repo
pipe.load_lora_weights("Ironictw2st/tk_3k_portraits",
                       weight_name="tk3k_portrait_v3.safetensors")

prompt = (
    "tk3k_portrait, solo, (full body:1.5), (full length:1.4), head to toe, "
    "standing, feet visible, man, lightly armored, "
    "flowing silk scholar robe, official's cap, dignified expression, "
    "full body portrait, (plain white background:1.4), simple background, "
    "isolated on white background"
)
negative = (
    "bust portrait, headshot, half body, cropped, close-up, waist up, "
    "multiple poses, reference sheet, multiple views, multiple characters, "
    "low quality, blurry, jpeg artifacts, signature, text, watermark, "
    "deformed, extra limbs, background, scenery, wall, weapon, sword"
)

image = pipe(
    prompt, negative_prompt=negative,
    num_inference_steps=30, guidance_scale=6.5,
    width=832, height=1216,                 # portrait aspect → full-body framing
    cross_attention_kwargs={"scale": 1.0},  # LoRA weight
).images[0]
image.save("portrait.png")

Prompting guide

Always start the prompt with the trigger: tk3k_portrait.
Then describe: man / woman → armor (unarmored / lightly armored / heavily armored) → free description (robes, cap, beard, expression, colors).
Full-body figures: add framing cues like solo, (full body:1.5), (full length:1.4), head to toe, standing, feet visible and render at 832×1216 (portrait). For a bust, drop those and use 1024×1024.
Clean background: end with full body portrait, (plain white background:1.4), simple background, isolated on white background. The model reliably produces a near-white background, which makes background removal easy.
Color, not literal elements: to evoke a Wuxing element, describe a palette (e.g. "deep red and gold robes") rather than tags like "fire" — element tags tend to render literal flames/water/VFX.

Recommended settings

Setting	Value
Sampler / scheduler	Euler Ancestral (`EulerAncestralDiscreteScheduler`)
Steps	30
CFG	6.5
LoRA weight	1.0
Resolution	832×1216 (full body) or 1024×1024 (bust)

Training

Base: Stable Diffusion XL 1.0
Data: portraits from the Total War: THREE KINGDOMS SAD modpack, composited on white
Method: LoRA fine-tune (kohya / sd-scripts). v3 is the final checkpoint.

License

Released under Creative Commons Attribution–NonCommercial 4.0 (CC BY-NC 4.0).

✅ Share and use for personal, non-commercial purposes, with attribution.
❌ No commercial use.
This adapter also requires and runs on SDXL 1.0, whose CreativeML Open RAIL++-M use-based restrictions also apply.
The underlying art style derives from Total War: THREE KINGDOMS and the SAD modpack; this is unofficial fan content and not affiliated with or endorsed by their owners.

Credits

Base model: Stability AI — Stable Diffusion XL 1.0
App & tooling: https://github.com/Ironictw2st/tkart-portrait-generator

Downloads last month: 3

Model tree for Ironictw2st/tk_3k_portraits

Base model

stabilityai/stable-diffusion-xl-base-1.0

Adapter

(9690)

this model

TW3K Portrait LoRA (tk3k_portrait)