mudler commited on
Commit
882bc04
·
verified ·
1 Parent(s): d580b23

Add Depth Anything V2 models to card

Browse files
Files changed (1) hide show
  1. README.md +43 -0
README.md CHANGED
@@ -5,6 +5,7 @@ pipeline_tag: depth-estimation
5
  tags:
6
  - depth-anything
7
  - depth-anything-3
 
8
  - depth-estimation
9
  - monocular-depth
10
  - camera-pose
@@ -20,6 +21,15 @@ base_model:
20
  - depth-anything/DA3MONO-LARGE
21
  - depth-anything/DA3METRIC-LARGE
22
  - depth-anything/DA3NESTED-GIANT-LARGE
 
 
 
 
 
 
 
 
 
23
  ---
24
 
25
  # Depth Anything 3 — GGUF weights for [depth-anything.cpp](https://github.com/mudler/depth-anything.cpp)
@@ -59,6 +69,39 @@ baked into the file; the loader reads them, nothing is hardcoded.
59
  > metric (ViT-L) branch together and aligns them to produce metric-scale depth + pose. Download
60
  > both `depth-anything-nested-anyview.gguf` and `depth-anything-nested-metric.gguf`.
61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
62
  ### Which one should I use?
63
 
64
  - **Just trying it out / CPU:** `depth-anything-base-q4_k.gguf` (99 MB, near-lossless).
 
5
  tags:
6
  - depth-anything
7
  - depth-anything-3
8
+ - depth-anything-2
9
  - depth-estimation
10
  - monocular-depth
11
  - camera-pose
 
21
  - depth-anything/DA3MONO-LARGE
22
  - depth-anything/DA3METRIC-LARGE
23
  - depth-anything/DA3NESTED-GIANT-LARGE
24
+ - depth-anything/Depth-Anything-V2-Small
25
+ - depth-anything/Depth-Anything-V2-Base
26
+ - depth-anything/Depth-Anything-V2-Large
27
+ - depth-anything/Depth-Anything-V2-Metric-Hypersim-Small
28
+ - depth-anything/Depth-Anything-V2-Metric-Hypersim-Base
29
+ - depth-anything/Depth-Anything-V2-Metric-Hypersim-Large
30
+ - depth-anything/Depth-Anything-V2-Metric-VKITTI-Small
31
+ - depth-anything/Depth-Anything-V2-Metric-VKITTI-Base
32
+ - depth-anything/Depth-Anything-V2-Metric-VKITTI-Large
33
  ---
34
 
35
  # Depth Anything 3 — GGUF weights for [depth-anything.cpp](https://github.com/mudler/depth-anything.cpp)
 
69
  > metric (ViT-L) branch together and aligns them to produce metric-scale depth + pose. Download
70
  > both `depth-anything-nested-anyview.gguf` and `depth-anything-nested-metric.gguf`.
71
 
72
+ ### Depth Anything V2
73
+
74
+ The same engine also runs [Depth Anything **V2**](https://github.com/DepthAnything/Depth-Anything-V2)
75
+ checkpoints. DA2 is **depth only** — no confidence, pose or sky. **Relative** models output an inverse
76
+ depth map through a `ReLU` head; **metric** models output depth in **metres** through a
77
+ `Sigmoid × max_depth` head (`max_depth=20` for the indoor Hypersim variants, `max_depth=80` for the
78
+ outdoor VKITTI variants). The ViT-g (Giant) DA2 checkpoint is not shipped (its `Depth-Anything-V2-Giant`
79
+ HF repo is gated/unreleased).
80
+
81
+ Each model below ships in f32 plus f16 / q8_0 / q6_k / q5_k / q4_k quants (only the f32 + a representative
82
+ quant are listed for brevity; the full set is in `SHA256SUMS`).
83
+
84
+ | File | Source checkpoint | Backbone | Depth type | Output |
85
+ |------|-------------------|----------|-----------|--------|
86
+ | `depth-anything2-small-f32.gguf` | `Depth-Anything-V2-Small` | ViT-S | relative | inverse depth |
87
+ | `depth-anything2-small-q8_0.gguf` | `Depth-Anything-V2-Small` | ViT-S | relative | inverse depth (near-lossless) |
88
+ | `depth-anything2-base-f32.gguf` | `Depth-Anything-V2-Base` | ViT-B | relative | inverse depth |
89
+ | `depth-anything2-large-f32.gguf` | `Depth-Anything-V2-Large` | ViT-L | relative | inverse depth |
90
+ | `depth-anything2-large-q4_k.gguf` | `Depth-Anything-V2-Large` | ViT-L | relative | inverse depth (smallest) |
91
+ | `depth-anything2-metric-hypersim-small-f32.gguf` | `Depth-Anything-V2-Metric-Hypersim-Small` | ViT-S | **metric** (≤20 m, indoor) | depth in metres |
92
+ | `depth-anything2-metric-hypersim-base-f32.gguf` | `Depth-Anything-V2-Metric-Hypersim-Base` | ViT-B | **metric** (≤20 m, indoor) | depth in metres |
93
+ | `depth-anything2-metric-hypersim-large-f32.gguf` | `Depth-Anything-V2-Metric-Hypersim-Large` | ViT-L | **metric** (≤20 m, indoor) | depth in metres |
94
+ | `depth-anything2-metric-vkitti-small-f32.gguf` | `Depth-Anything-V2-Metric-VKITTI-Small` | ViT-S | **metric** (≤80 m, outdoor) | depth in metres |
95
+ | `depth-anything2-metric-vkitti-base-f32.gguf` | `Depth-Anything-V2-Metric-VKITTI-Base` | ViT-B | **metric** (≤80 m, outdoor) | depth in metres |
96
+ | `depth-anything2-metric-vkitti-large-f32.gguf` | `Depth-Anything-V2-Metric-VKITTI-Large` | ViT-L | **metric** (≤80 m, outdoor) | depth in metres |
97
+
98
+ **Parity.** Every DA2 GGUF is verified against the upstream `DepthAnythingV2` forward (correlation > 0.999
99
+ end-to-end at f32, q8_0 near-lossless at corr 0.99962, q4_k at 0.99944). The one exception is
100
+ `depth-anything2-metric-vkitti-small` at corr **0.9983** — this is **not a porting defect** (the C++ route
101
+ matches the reference `Sigmoid × 80` math exactly); it is the inherent ≤20× amplification of backbone
102
+ fp-rounding noise by the widest metric scale on the smallest backbone. Absolute error stays sub-1%
103
+ (mean 0.57% of 80 m), and the same ViT-S backbone scores 0.9996 in relative mode. Accepted as near-lossless.
104
+
105
  ### Which one should I use?
106
 
107
  - **Just trying it out / CPU:** `depth-anything-base-q4_k.gguf` (99 MB, near-lossless).