Besch Style - Flux.2 Klein-9B Base LoKR

Cleaner silhouette separation

Slightly more composed and controlled close-ups

V02 is technically more consistent and responds better to structured prompts.

Captioning Philosophy (V02)

For V02, I rebuilt the dataset captions with a structured taxonomy approach:

1️⃣ Style Layer

Consistent use of controlled stylistic vocabulary:

bold dark ink outlines

flat graphic illustration

layered painterly planes

graphic high contrast mode

angular shadow segmentation

warm-to-cool rim lighting

2️⃣ Structural Layer

Clear separation between:

subject description

camera framing

lighting

environment

texture treatment

3️⃣ Vocabulary Normalization

Synonyms were reduced and harmonized. Redundant adjectives removed. Key tokens stabilized.

The goal was to reduce noise in the training signal and give the LoKR a clearer stylistic identity.

This made V02 more controllable — but slightly less raw in emotional close-ups compared to V01.

Training Pipeline

Base: Klein-9B

Adapter: LoKR

Framework: Diffusers

Trainer: SimpleTuner

Inference: ComfyUI

The model was trained using SimpleTuner , a modular Diffusers-based trainer that allows precise control over:

learning rate scheduling

optimizer configuration

dataset structure

LoRA/LoKR behavior

resolution strategy

The training pipeline was kept clean and reproducible. No merges, no hidden tricks.

Resolution Notes

The dataset contains a mix of 512, 768 and some 1024 images.

Close-ups tend to perform best around:

768–1024 base resolution

Very high initial resolutions (1MP x1.25 and above) may slightly destabilize hand placement and complex limb overlap due to spatial distribution differences in training data.