maximsobolev275/BFS-Qwen21-Best-Face-Swap-Fix

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:

1. Base Model (No LoRA applied)
2. Original BFS LoRA (Notice the blurriness and smoothed skin texture)
3. Optimized Variants (Pruned / DARE drop-masked versions restoring sharp details)

Comparison — Seed 0

Comparison — Seed 1

Comparison — Seed 2

Comparison — Seed 3