Instructions to use wardner/Depth-Anything-V3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- RKLLM
How to use wardner/Depth-Anything-V3 with RKLLM:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Download DepthAnythingV3.json from wardner/Depth-Anything-V3: direct link, hf CLI and curl.
- Browser
- Download file 1.94 kB
-
https://huggingface.co/wardner/Depth-Anything-V3/resolve/main/DepthAnythingV3.json
- Command line
-
hf download hf://wardner/Depth-Anything-V3/DepthAnythingV3.json
-
curl -L -o DepthAnythingV3.json https://huggingface.co/wardner/Depth-Anything-V3/resolve/main/DepthAnythingV3.json
1.94 kB
| { | |
| "label": "[Depth]视觉深度信息v3", | |
| "capability_types": [ | |
| "swcap_vision_depth" | |
| ], | |
| "spec_factory": "depth_anything_v3", | |
| "worker_name": "depth-v3", | |
| "model_dir_name": "Depth-Anything-V3", | |
| "model_dir_env": "WDNAI_DEPTH_ANYTHING_V3_DIR", | |
| "instance_name": "depth-v3", | |
| "memory_estimate_mb": 3000, | |
| "estimated_init_seconds": 30, | |
| "description": "视觉深度信息v3,给定图像输出逐像素深度图与置信度(Depth Anything V3 单目深度估计)", | |
| "notes": "【当前部署】1-view / 280×280 导出(global 段输入 [1,401,1,768]、head 段输出 [1,1,280,280])—— 一张图即可出深度图 + 置信度。view 数是**导出时**参数(global 段输入形状 [V,L,1,D],不是运行期参数),换 V 只需重新导出三段 RKNN(model-zoo 的 --views,已验证范围 1-10),**SDK 与本服务无需改动**。部署侧文件命名跟 LocateAnything-3B 一致:<role>_DepthAnythingV3.280x280.1view.1828.rknn + 同名 .weight(去掉 .1828 段)—— 共 6 个文件。多相机几何另需 model-zoo 导出 ray 分支(当前导出脚本只出 depth+depth_conf,不支持)并换更大的模型——论文数据显示 DA3-Base 的位姿精度仅为 DA3-Giant 的 1/4,位姿比深度更吃模型规模。单目深度估计本身(kind=\"depth\",图像 → 深度图 + 置信度,不是文本)。RKNN 导出被拆成三段:local(逐图 token)→ global(跨 view 融合)→ head(出深度/置信度),三段**共享**一份按 NPU 核分配的内部显存(C++ demo 的 da3_init_internal_mem_share,约 1.6GB),少了这一步 model_init 会 MODEL_SETUP fail。权重目录由 wdnai 内部的 _default_depth_anything_v3_paths 解析,可用 WDNAI_DEPTH_ANYTHING_V3_DIR 覆写。模型未部署时 load 会抛 FileNotFoundError,由 worker 的 load 流程转成 errored 状态。", | |
| "available_targets": [ | |
| "182x" | |
| ], | |
| "display_order": 80 | |
| } | |