Skip to content

v1.2 — Perceptual + temporal fine-tune

Latest

Choose a tag to compare

@crispianm crispianm released this 21 Jul 11:41

dabit_perceptual.pth — the v1.1 dabit_retrained.pth model fine-tuned for 50k iterations with an added LPIPS-VGG perceptual loss and a flow-warped, occlusion-masked temporal-consistency loss (Lai et al., ECCV 2018). Config: configs/dabit_ft.json, losses.perc_weight / losses.temp_weight.

Result (all 90 DAVIS-Blur sequences, fp32)

Model PSNR ↑ SSIM ↑ LPIPS ↓ tOF ↓
v1.1 retrain 29.17 0.858 0.217 1.103
v1.2 perceptual 28.99 0.856 0.173 0.882

Trades ~0.18 dB PSNR for 20% lower LPIPS (sharper perceptual detail) and 20% lower tOF (smoother motion). Choose this model for perceptual quality, the v1.1 retrain for peak PSNR/SSIM.

Install

Requires the RAFT + Depth Anything V2 dependency weights from v1.1:

mkdir -p weights
gh release download v1.2 -R crispianm/DaBiT -D weights   # perceptual model
gh release download v1.1 -R crispianm/DaBiT -D weights   # deps (raft, depth-anything)
python test_dabit.py --model weights/dabit_perceptual.pth