## French Manicure V2 LoRA
Updated version of v1 trained with a cleaner dataset. Slightly improved performance (you can use heavier weight).
A LoRA fine-tune of Flux.1 D trained on a curated dataset of french manicure v2 images.
## Usage
Load this LoRA in ComfyUI using LoraLoaderModelOnly (trained on unet only) with strength 0.8u20131.0.
## Training Details
| Parameter | Value |
|-----------|-------|
| Base model | Flux.1 D |
| Network rank (dim) | 16 |
| Network alpha | 1.0 |
| Precision | bf16 |
| Steps | 1500 |
| Learning rate | 0.0004 |
| Optimizer | CAME |
| LR scheduler | constant |
| Training resolutions | mixed |
| Bucket training | No u2014 see dataset config |
| Weighting scheme | logit_normal |
| Timestep sampling | shift |
## Compatibility
- Base model: Flux.1 D
- Node: LoraLoaderModelOnly (unet-only u2014 no text encoder weights)
- Recommended strength: 0.8 u2013 1.0
- Sampler: Euler + Simple scheduler, 20 steps
- Guidance: FluxGuidance 3.5 - 10 (for Flux)
Description
Trained 1500 steps on Flux.1 D. Rank 16, alpha 1.0, bf16 precision. Optimizer: CAME, LR: 0.0004.