kn1febrush-flux-style-lora

kn1febrush-flux-style-lora

Style LoRA for FLUX.1-dev. Trained on 14 curated reference images.

Trigger: kn1febrush

Usage (ComfyUI)

UNETLoader (flux1-dev) → LoraLoaderModelOnly → KSampler. Flux dev is CFG-distilled: keep cfg = 1.0 and drive adherence with FluxGuidance (~3.0).

Prompt template:

kn1febrush, a full body character illustration of <SUBJECT>, <STYLE CUES>

Strength guide — 0.6 lets the base model's smooth prior show through; 0.8–1.0 is the usable range; above 1.0 training subjects and poses start reappearing.

Contents

• *.safetensors — final weights plus intermediate checkpoints. Style LoRAs overfit, so the best checkpoint is often not the last one. Compare them.
• dataset/ — training images and caption files, so the run is reproducible.
• samples/ — outputs per checkpoint with the generation settings recorded.

Training

• Base: FLUX.1-dev bf16, kohya sd-scripts flux_train_network.py
• LoRA dim 24 / alpha 24, unet only
• adamw8bit, lr 1e-4, constant_with_warmup (40 warmup steps)
• 2000 steps, batch 1, --fp8_base, gradient checkpointing
• timestep_sampling shift, discrete_flow_shift 3.1582, model_prediction_type raw, guidance_scale 1.0
• Aspect-ratio bucketing, 1024 base resolution

Captioning method

Captions describe content only — who the character is, what they wear, what they do. No style words at all. Everything constant across the set and absent from the captions collapses onto the shared trigger token, which is what makes the style strong and strength-controllable. Writing style words into captions scatters it across those common words and produces a weak LoRA.