Improved Depth Anything V2 - Base model

[Jokeru18]

Ultimate DAV2 ViT-B — ultimate_dav2_vitB_INet_quality.safetensors

A drop-in ViT-B depth model distilled from DAV2-L @518. It beats the official DAV2 ViT-B on pairwise accuracy (97.39% vs 97.34%) and cuts the student→teacher error by up to 74% on DA-2K — while keeping the same architecture, the same input normalization, and the same API.

What is this

• Architecture: Depth Anything V2, ViT-B (12 blocks), arbitrary-resolution input (trained @518)
• Input normalization: ImageNet (AR)
• Teacher: DAV2-L @518
• Run: ultimate_LAB_DAV2 / VitB_ImageNet_AR → ultimate_dav2_vitB_INet_quality.safetensors
• Format: .safetensors — drop-in replacement for depth_anything_v2_vitb

Headline results

Benchmark	Metric	Official DAV2 ViT-B	ultimate_dav2_vitB_INet_quality.safetensors	Δ (relative)
DA-2K vs Teacher (1030 img)	AbsMed error	0.17579	0.04474	−74.5%
DA-2K vs Teacher (1030 img)	AbsP95 error	0.50268	0.21525	−57.2%
DA-2K vs Teacher (1030 img)	MID error	0.17723	0.05676	−68.0%
DA-2K vs Teacher (1030 img)	FAR error	0.17117	0.04479	−73.8%
DA-2K vs Teacher (1030 img)	Ordinal acc.	0.9319	0.9771	+4.8%
DA-2K vs Teacher (1030 img)	SSIM	0.8014	0.9374	+17.0%
DA-2K pairwise (2068 pairs)	Accuracy	97.34%	97.39%	+0.05 pt
NYUv2 vs GT (654 img)	AbsMed error	0.21029	0.16114	−23.4%
NYUv2 vs GT (654 img)	MID error	0.23642	0.19242	−18.6%
NYUv2 vs GT (654 img)	FAR error	0.24050	0.17111	−28.9%
NYUv2 vs GT (654 img)	SSIM	0.7280	0.7736	+6.3%

Benchmark 1 — DA-2K, fidelity to Teacher (DAV2-L @518)

1030 images, student canvas 518, teacher computed once per image and cached.

Metric	Official ViT-B	ultimate_dav2_vitB_INet_quality.safetensors
Spearman ρ	0.9924	0.9912
AbsMed	0.17579	0.04474
AbsP95	0.50268	0.21525
Ordinal	0.9319	0.9771
SSIM	0.8014	0.9374
NEAR err	0.30315	0.12370
MID err	0.17723	0.05676
FAR err	0.17117	0.04479

The student→teacher gap is essentially closed on absolute error: what the official ViT-B loses against the L teacher, ultimate_dav2_vitB_INet_quality.safetensors loses only a fraction of (median error ~4x lower).

Benchmark 2 — NYUv2, against saved ground truth

654 images, GT .dat depth maps re-aligned to each canvas.

Metric	Official ViT-B	ultimate_dav2_vitB_INet_quality.safetensors
Spearman ρ	0.9483	0.9455
AbsMed	0.21029	0.16114
AbsP95	0.74138	0.69535
Ordinal	0.8461	0.8546
SSIM	0.7280	0.7736
NEAR err	0.33934	0.33398
MID err	0.23642	0.19242
FAR err	0.24050	0.17111

Largest wins: FAR-range error −28.9%, MID −18.6%, AbsMed −23.4%.

Benchmark 3 — DA-2K pairwise accuracy (1033 images, 2068 pairs)

Scene	Official ViT-B	ultimate_dav2_vitB_INet_quality.safetensors
adverse_style	96.95%	96.65%
aerial	99.48%	98.97%
indoor	97.38%	97.14%
non_real	99.01%	98.68%
object	97.97%	97.30%
outdoor	95.64%	96.22%
transparent_reflective	95.79%	96.73%
underwater	97.44%	99.15%
Overall	97.34%	97.39%

Overall accuracy is higher, and the gains come exactly where the official ViT-B is weakest: outdoor, transparent/reflective, underwater.

Low-resolution robustness

The teacher-fidelity advantage survives downscaling — at the lowest tested canvas (294) on DA-2K:

Metric @294	Official ViT-B	ultimate_dav2_vitB_INet_quality.safetensors
AbsMed	0.17660	0.06653 (−62.3%)
AbsP95	0.56160	0.33102 (−41.1%)

The official ViT-B degrades 2.4× in AbsMed from 518→294; ultimate_dav2_vitB_INet_quality.safetensors degrades only 1.5×.

Honest notes

• Spearman ρ is marginally lower (−0.0012 DA-2K, −0.0028 NYUv2) — within noise, but not claimed as a win.
• Small per-scene losses on adverse_style / aerial / object in pairwise (≤0.7 pt each), offset by larger gains elsewhere.
• Benchmarks: DA-2K 1030–1033 images; NYUv2 654 images; all metrics at native 518 canvas unless stated.

Usage

Drop-in replacement for depth_anything_v2_vitb — same encoder, same ImageNet normalization, same AR forward pass:

# point your loader at ultimate_dav2_vitB_INet_quality.safetensors
# (any DAV2 ViT-B pipeline accepts it as-is)

Reproducibility

• Teacher: DAV2-L @518 · Student: DAV2 ViT-B, ImageNet AR, 12 blocks

##Instalation : Copy the files and overwrite over your Nunif folder.
Optionally You can download manually the Safetensors form the repository. Otherwise they will be downloaded automatically on first start.

nunif.zip