AZIT · LoRA Pipeline
LoRA Training Pipeline Manual
FLUX.2 Klein 9B-based AI Artist LoRA Training · RTX 5090 (32GB) · AI-Toolkit · ComfyUI · ArcFace
0.732
Male actor PEAK
Step 2200
0.692
Female actor PEAK
Step 1500
CFG 3.0
Optimal CFG
Steps 70 · Shift 0.5
0.6050
LoRA vs Dataset
Beats HF 0.5090
Pipeline Flow
1 Data Preparation
- Dataset bible (principles)
- Training strategy (reference)
- Image preprocessing (crop / normalize)
- Quality checks (resolution / tone / composition)
2 LoRA Training
- AI-Toolkit training config
- ComfyUI API generation requests
- Batch training run
- Live overfit monitoring
3 Quality Evaluation
- ArcFace per-step measurement
- Higgsfield benchmark comparison
- PEAK / decline pattern analysis
- Best checkpoint selection
4 Parameter Optimization
- CFG sweep (1.0–7.0)
- Steps sweep (20–100)
- ModelSamplingFlux sweep
- Automated ArcFace analysis
Outputs
.safetensors LoRA weight files · Pipeline execution report · ArcFace benchmark results
Headline Results
| Metric | Value |
|---|
| Base model | FLUX.2 Klein 9B |
| Best checkpoint (female actor) | Step 1500 (AVG 0.692) |
| Best checkpoint (male actor) | Step 2200 (AVG 0.732) |
| ArcFace LoRA vs Dataset | 0.6050 (beats HF 0.5090) |
| Optimal CFG | 3.0 |
| Optimal Sampling Steps | 70 |
| Optimal max_shift | 0.5 |
File Structure
H
00_pipeline_manual
This file
M
01_dataset_bible
Dataset principles
M
02_training_strategy_djdante
Klein 9B strategy
T
03_arcface_benchmark_results
Experiment log
T
04_pipeline_execution_report
Run results
Y
05_training_config_female_actor
FLUX config
Y
06_training_config_male_actor
FLUX config
H
07_dataset_guide_web_document
Visualization document
scripts/
PY
3_ComfyUI_workflow_dispatch
PY
5_ArcFace_per_step_measurement
PY
6_benchmark_comparative_analysis
PY
8_sweep_results_analysis