BFS (Best Face Swap) — Optimized & De-blurred LoRA Variants
This repository contains cleaned and optimized variations of the Alissonerdx/BFS-Best-Face-Swap LoRA.
Motivation & Problem
While the original BFS LoRA provides solid face swap transfer, it often introduces significant blurriness, fine texture loss, and a plastic/soapy degradation across skin areas.
To resolve this issue while retaining identity likeness, these variants were optimized using weight pruning and stochastic drop masking:
- Pruning (20%): Removes the bottom 20% weakest/noisy weight layers that cause unwanted model drift and texture degradation.
- DARE Drop Masking (without Rescaling): Random drop-masking (density 0.50–0.60) applied without magnitude rescaling. Skipping the rescaling step prevents artificial weight amplification, successfully stripping away soapy blur while keeping the facial structure intact.
Available Variants
| File | Pruning & Masking | Description |
|---|---|---|
bfs_dare60-avgrank30.safetensors |
Drop Mask (density 0.60, no rescale), avg rank 30 | Gentle cleanup; preserves strong likeness while noticeably reducing original blur. |
bfs_prune20_dare50-avgrank26.safetensors |
20% Pruned + Drop Mask (density 0.50, no rescale), avg rank 26 | Strongest de-blurring and detail retention; best for crisp skin textures and sharp eyes. |
bfs_prune20_dare60-avgrank30.safetensors |
20% Pruned + Drop Mask (density 0.60, no rescale), avg rank 30 | Balanced profile: removes muddy blur while keeping likeness very close to the original. |
Reference Inputs
Reference Head (ref_head.png) |
Reference Body (ref_body.png) |
|---|---|
Visual Comparisons Across Seeds
The comparison grids below demonstrate:
- Base Model (No LoRA applied)
- Original BFS LoRA (Notice the blurriness and smoothed skin texture)
- Optimized Variants (Pruned / DARE drop-masked versions restoring sharp details)