jayark/pfa-lora

PFA Style LoRA

A style LoRA trained on Qwen-Image-2512 that captures the Proud Family: Louder and Prouder art style. Generate any subject — new characters, scenes, environments — rendered in the show's distinctive look.

Trigger Word

pfa style

Prepend pfa style, to the beginning of every prompt. This is the token that activates the style.

Prompting Guide

Use natural language, not tags

Qwen-Image-2512 uses Qwen2.5-VL as its text encoder, which expects natural language sentences — not booru-style comma-separated tags.

Good:

pfa style, A close-up front-facing portrait of a teenage girl smiling warmly
in a school hallway with warm interior lighting.

Bad:

pfa style, 1girl, close-up, portrait, front view, teenager, smile, school hallway

What to describe in your prompt

• Subject identity (age, skin tone, hair, clothing, expression)
• Action or pose
• Setting and environment details
• Lighting and time of day
• Camera angle and composition (close-up, wide shot, full-body, etc.)

What NOT to put in your prompt

The trigger word carries the visual style. Do not describe the art style itself:

• Avoid: cartoon, animated, 2D, drawn, illustrated, stylized, anime
• Avoid: art-medium or quality words

Example prompts

pfa style, A close-up front-facing portrait of a teenage girl smiling warmly
in a school hallway with warm interior lighting.

pfa style, A wide shot of a suburban house and front yard under a bright sunny
blue sky, no people in frame.

pfa style, A full-body shot of a businessman in a gray suit holding a coffee cup,
standing at a city bus stop in the rain under an overcast sky.

pfa style, A teenage girl with braids reading on her bed in a cozy bedroom
with fairy lights.

pfa style, A wide suburban street at golden hour with a black family walking
their dog.

Checkpoints

Three checkpoints are provided at different training stages:

File	Epochs	Notes
pfa-style-v4-200.safetensors	200	Early style emergence
pfa-style-v4-400.safetensors	400	Expected sweet spot
pfa-style-v4-600.safetensors	600	Strongest style, watch for overcook

Lower epoch checkpoints preserve more subject detail; higher epochs push harder toward the show's look. Try 400 first.

Inference Parameters

Parameter	Value
Inference steps	20
Guidance scale	4.0
LoRA multiplier	1.0 (reduce to 0.75 if overcooked)
Image size	832x832, 480x832, or 832x480
Attention mode	torch

Recommended negative prompt

photorealistic, hyperrealistic, flat white eyes, no iris, dot pupils, creepy eyes

Training Details

• Base model: Qwen-Image-2512
• Method: LoRA via musubi-tuner
• Rank: 16
• Alpha: 16
• Optimizer: AdamW 8-bit
• Learning rate: 5e-5
• Mixed precision: bf16
• Dataset: 30 images with natural-language captions