Flux-2K-9B-32-05 LoRA
Style LoRA trained on FLUX.2-klein-base-9B.
Details
- Base model:
black-forest-labs/FLUX.2-klein-base-9B
- Architecture: flux2_klein_9b
- Trigger word:
t01nspc_style
- Network: LoRA (linear=32, conv=16)
- Training steps: 9000
- Learning rate: 5e-5
- Optimizer: AdamW 8bit
- Dtype: bf16
- Dataset: 222 images (222_ds) with danbooru-style captions
- Resolution: 512, 768, 1024
- Checkpoint interval: every 500 steps
- Content/Style balance: balanced
- Diff output preservation: enabled (class: photo)
Training Parameters
| Parameter | Value |
|---|
| Batch size | 1 |
| Gradient accumulation | 1 |
| Noise scheduler | flowmatch |
| Timestep type | weighted |
| Loss type | MSE |
| Caption dropout | 5% |
| Weight decay | 0.00015 |
| Gradient checkpointing | ✓ |
Checkpoints
12 checkpoints available in checkpoints/ folder (steps 1000-6500, every 500 steps).
Usage
Use with FLUX.2-klein-9B compatible pipelines. Add t01nspc_style to your prompt to activate the style.
