<h1 id="besch-style-flux2-klein-9b-lokr">BESCH Style – Flux2 / Klein-9B LoKR</h1><h2 id="overview">Overview</h2>BESCH is a graphic illustration style focused on:<ul><li>bold dark ink outlines with variable line weight</li><li>layered painterly planes</li><li>faceted hair masses with unified flow</li><li>warm-to-cool edge lighting</li><li>high contrast graphic rendering</li><li>expressive close-up intensity</li><li>selective halftone or texture accents</li></ul>The goal of this style is to balance strength and softness, controlled structure and emotional intensity.It performs especially well in:<ul><li>close-up portraits</li><li>dynamic three-quarter views</li><li>graphic hero compositions</li><li>stylized color blocking</li></ul><h1 id="versions">Versions</h1><h2 id="v01-expressive-base">🔴 V01 – Expressive Base</h2><ul><li>No trigger required</li><li>Higher learning rate (5e-5)</li><li>Same image dataset as V02</li><li>Less structured caption organization</li></ul>Characteristics:<ul><li>More expressive close-ups</li><li>Stronger emotional intensity</li><li>Slightly looser structure</li><li>Sometimes less precise in full-body compositions</li></ul>V01 tends to push facial expression and mood more aggressively.<hr /><h2 id="v02-structured-caption-edition">🔵 V02 – Structured Caption Edition</h2>Trigger: <code>besch</code><ul><li>Lower learning rate (3e-5)</li><li>Same image dataset as V01</li><li>Fully reorganized captions</li><li>Vocabulary normalization</li><li>Structured tagging and stylistic taxonomy</li></ul>Characteristics:<ul><li>More stable full-body rendering</li><li>Better control of graphic elements</li><li>Cleaner silhouette separation</li><li>Slightly more composed and controlled close-ups</li></ul>V02 is technically more consistent and responds better to structured prompts.<hr /><h1 id="captioning-philosophy-(v02)">Captioning Philosophy (V02)</h1>For V02, I rebuilt the dataset captions with a structured taxonomy approach:<h3 id="1-style-layer">1️⃣ Style Layer</h3>Consistent use of controlled stylistic vocabulary:<ul><li>bold dark ink outlines</li><li>flat graphic illustration</li><li>layered painterly planes</li><li>graphic high contrast mode</li><li>angular shadow segmentation</li><li>warm-to-cool rim lighting</li></ul><h3 id="2-structural-layer">2️⃣ Structural Layer</h3>Clear separation between:<ul><li>subject description</li><li>camera framing</li><li>lighting</li><li>environment</li><li>texture treatment</li></ul><h3 id="3-vocabulary-normalization">3️⃣ Vocabulary Normalization</h3>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.<hr /><h1 id="training-pipeline">Training Pipeline</h1><ul><li>Base: Klein-9B</li><li>Adapter: LoKR</li><li>Framework: Diffusers</li><li>Trainer: SimpleTuner</li><li>Inference: ComfyUI</li></ul>The model was trained using SimpleTuner, a modular Diffusers-based trainer that allows precise control over:<ul><li>learning rate scheduling</li><li>optimizer configuration</li><li>dataset structure</li><li>LoRA/LoKR behavior</li><li>resolution strategy</li></ul>The training pipeline was kept clean and reproducible. No merges, no hidden tricks.<hr /><h1 id="resolution-notes">Resolution Notes</h1>The dataset contains a mix of 512, 768 and some 1024 images.Close-ups tend to perform best around:<ul><li>768–1024 base resolution</li></ul>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.<hr /><h1 id="stacking-recommendation">Stacking Recommendation</h1>Stacking V01 + V02 can produce interesting results:<ul><li>V01 at lower weight for emotional intensity</li><li>V02 as structural backbone</li></ul>Example:<ul><li>V02: 0.7</li><li>V01: 0.3–0.4</li></ul>This often restores expressiveness while keeping compositional control.<hr /><h1 id="prompting-tips">Prompting Tips</h1>Works well with:<ul><li>dynamic wind interaction in hair</li><li>three-quarter perspective</li><li>off-center framing</li><li>edge lighting emphasis</li><li>graphic shadow segmentation</li></ul>Avoid overloading with random style synonyms. V02 responds best to controlled vocabulary.<hr /><h1 id="artistic-intent">Artistic Intent</h1>BESCH is not meant to be hyper-realistic.It lives in the tension between:<ul><li>graphic clarity</li><li>painterly layering</li><li>emotional intensity</li><li>controlled stylization</li></ul>Strength and tenderness can coexist in the same frame.

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

Structured Caption EditionTrigger: <code>besch</code><ul><li>Lower learning rate (3e-5)</li><li>Same image dataset as V01</li><li>Fully reorganized captions</li><li>Vocabulary normalization</li><li>Structured tagging and stylistic taxonomy</li></ul>Characteristics:<ul><li>More stable full-body rendering</li><li>Better control of graphic elements</li><li>Cleaner silhouette separation</li><li>Slightly more composed and controlled close-ups</li></ul>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 LayerConsistent use of controlled stylistic vocabulary:<ul><li>bold dark ink outlines</li><li>flat graphic illustration</li><li>layered painterly planes</li><li>graphic high contrast mode</li><li>angular shadow segmentation</li><li>warm-to-cool rim lighting</li></ul>2️⃣ Structural LayerClear separation between:<ul><li>subject description</li><li>camera framing</li><li>lighting</li><li>environment</li><li>texture treatment</li></ul>3️⃣ Vocabulary NormalizationSynonyms 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<ul><li>Base: Klein-9B</li><li>Adapter: LoKR</li><li>Framework: Diffusers</li><li>Trainer: SimpleTuner</li><li>Inference: ComfyUI</li></ul>The model was trained using SimpleTuner, a modular Diffusers-based trainer that allows precise control over:<ul><li>learning rate scheduling</li><li>optimizer configuration</li><li>dataset structure</li><li>LoRA/LoKR behavior</li><li>resolution strategy</li></ul>The training pipeline was kept clean and reproducible. No merges, no hidden tricks. Resolution NotesThe dataset contains a mix of 512, 768 and some 1024 images.Close-ups tend to perform best around:<ul><li>768–1024 base resolution</li></ul>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.

Files with SHA256: 13f115bf33f98cb5b71a9613b7435221693486a28a92ffb8cf9e6c99280d71b8