BESCH Style – Flux2 / Klein-9B LoKR

Overview

BESCH is a graphic illustration style focused on:

• bold dark ink outlines with variable line weight

• layered painterly planes

• faceted hair masses with unified flow

• warm-to-cool edge lighting

• high contrast graphic rendering

• expressive close-up intensity

• selective halftone or texture accents

The goal of this style is to balance strength and softness, controlled structure and emotional intensity.

It performs especially well in:

• close-up portraits

• dynamic three-quarter views

• graphic hero compositions

• stylized color blocking

Versions

🔴 V01 – Expressive Base

• No trigger required

• Higher learning rate (5e-5)

• Same image dataset as V02

• Less structured caption organization

Characteristics:

• More expressive close-ups

• Stronger emotional intensity

• Slightly looser structure

• Sometimes less precise in full-body compositions

V01 tends to push facial expression and mood more aggressively.

🔵 V02 – Structured Caption Edition

Trigger: besch

• Lower learning rate (3e-5)

• Same image dataset as V01

• Fully reorganized captions

• Vocabulary normalization

• Structured tagging and stylistic taxonomy

Characteristics:

• More stable full-body rendering

• Better control of graphic elements

• 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.

Stacking Recommendation

Stacking V01 + V02 can produce interesting results:

• V01 at lower weight for emotional intensity

• V02 as structural backbone

Example:

• V02: 0.7

• V01: 0.3–0.4

This often restores expressiveness while keeping compositional control.

Prompting Tips

Works well with:

• dynamic wind interaction in hair

• three-quarter perspective

• off-center framing

• edge lighting emphasis

• graphic shadow segmentation

Avoid overloading with random style synonyms.
V02 responds best to controlled vocabulary.

Artistic Intent

BESCH is not meant to be hyper-realistic.

It lives in the tension between:

• graphic clarity

• painterly layering

• emotional intensity

• controlled stylization

Strength and tenderness can coexist in the same frame.