Lawb — FLUX.2 Klein 4B
Character LoRA for Lawb. Trigger word: Lawb.
This is a
FLUX.2 retrain, not a conversion. The original
noema-art/Lawb is a FLUX.1-dev LoRA and remains the
artifact to use on FLUX.1 — FLUX.1 LoRAs do not apply to FLUX.2.
Usage
Base model: black-forest-labs/FLUX.2-klein-base-4B. Put Lawb in the prompt.
Holds identity best when the character is medium-to-close in frame. At small-in-frame, photoreal
framing the character can degrade toward a human silhouette — prompt the subject larger, or raise the
LoRA weight, if that shows up.
Training
| |
|---|
| Base | black-forest-labs/FLUX.2-klein-base-4B (arch flux2_klein_4b) |
| Rank / alpha | 32 / 32 |
| Steps | 4000 |
| LR | 1e-4, adamw8bit, weight_decay 1e-5 |
| Scheduler | flowmatch, timestep_type: weighted |
| Batch | 1, gradient_accumulation 1 |
| EMA | on, decay 0.99 |
| Resolution | 512 / 768 / 1024 (bucketed) |
| Caption dropout | 0.05 |
| Dataset | 30 images + caption sidecars |
| Trained with | ai-toolkit |
Hyperparameters match the original FLUX.1 run where they carry over (rank/alpha, steps, LR,
optimizer, scheduler, batch, caption dropout, resolution, trigger, EMA). The base model differs, and
the captions were regenerated — the originals were not recoverable — so this is not a byte-for-byte
reproduction of the FLUX.1 LoRA.
An otherwise identical no-EMA run was trained and compared on a fixed-seed five-prompt grid; EMA was
better on every prompt and is what ships here.
