NOTICE

Prompt-Depth-Anything-Small -- unofficial transformers-format package (trust_remote_code)
Packaged by bigshanedogg, 2026. NOT an official release of the original authors.

Components
----------
1. Prompt Depth Anything weights (model.ckpt, prompt-depth-anything-vits)
   Source: https://huggingface.co/depth-anything/prompt-depth-anything-vits
   Copyright: Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Jiaming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, Bingyi Kang
   License: Apache-2.0 -- LICENSE
2. Prompt Depth Anything code (promptda/promptda.py, model/, utils/logger.py; vendored into promptda_model.py; load_image / load_depth of utils/io_wrapper.py re-typed in image_processing_prompt_depth_anything.py)
   Source: https://github.com/DepthAnything/PromptDA @ 9161547ca7151e14fcc8f490ad3cc02fa9437483
   Copyright: the Prompt Depth Anything authors
   License: Apache-2.0 -- LICENSE
3. DINOv2 code (the repo's torchhub copy of facebookresearch/dinov2; vendored into promptda_dinov2.py)
   Source: https://github.com/DepthAnything/PromptDA/tree/main/torchhub/facebookresearch_dinov2_main @ 9161547ca7151e14fcc8f490ad3cc02fa9437483
   Copyright: Copyright (c) Meta Platforms, Inc. and affiliates.
   License: Apache-2.0 -- LICENSE.dinov2

Changes made in this package
----------------------------
- promptda_dinov2.py / promptda_model.py: the official files concatenated into flat files; package-internal imports replaced by relative imports; NestedTensorBlock spelled out where vision_transformer.py imported it as Block. No computation changed.
- configuration_ / modeling_ / image_processing_ / processing_prompt_depth_anything.py: new transformers wrappers; the model builds DINOv2 with the torchhub hub's arguments instead of torch.hub.load and runs PromptDA.forward; the image processor re-types io_wrapper.load_image / load_depth (that module imports open3d / matplotlib), checked bit-exact.
- model.safetensors: the Lightning model.ckpt state dict without its 'pipeline.' prefix (as PromptDA.load_checkpoint), converted by save_pretrained, every tensor checked equal; the training state (epoch, loops) dropped.

License files
-------------
- LICENSE: Apache License 2.0, copy from DepthAnything/PromptDA (code and weights)
- LICENSE.dinov2: Apache License 2.0, copy from the repo's torchhub/facebookresearch_dinov2_main