README.md

---
license: other
license_name: mixed-cc-by-nc-sa-4.0-and-mit
tags:
  - video-saliency
  - shot-boundary-detection
  - safetensors
  - reframe
library_name: safetensors
---

# reframe-asd-weights

Verified, re-containered **safetensors** weights used by the Reframe vertical-video
pipeline. Each file was produced offline in a trusted environment by unpickling the
original author checkpoint with `torch.load(..., weights_only=True)`, flattening to a
plain `{str: Tensor}` state-dict, saving via `safetensors.torch.save_file`, then
**reloading and proving exact tensor-equality** (`shape` + `dtype` + `torch.equal`) on
every tensor against the original. No tensor values were altered — these are pure
re-containers of the upstream weights. The `sha256` values below pin the **hosted
safetensors bytes**.

> Licenses differ per file (see each section). This repository is redistributed for
> **personal / non-commercial** use. The ViNet-S weight is CC BY-NC-SA 4.0 (ShareAlike);
> the TransNetV2 weight is MIT.

---

## `vinet-s-saliency.safetensors` — ViNet-S video saliency (DHF1K)

- **File size:** 38,138,372 bytes
- **sha256:** `803e6d265d46d3f4f3d7ec2c6c2f3b4511f9ba176aa12e348ac317788ca0dc68`
- **Tensors:** 470, exact tensor-equality vs original VERIFIED
- **Original source:** author checkpoints bundle on Google Drive
  (file id `12UeAsdiD2xPLmoLRDcE_HjAUjxFdmw5N`, `checkpoints.tar.gz`, ~2.83 GiB),
  member `final_models/ViNet_S/vinet_s_visual_dataset_models/vinet_s_dhf1k.pt`
  (38,266,829 bytes, sha256 `5d097a6b145b2cff7f08aa141a91e7aec4ac967504b439f4b04110c7e475cbbd`).
- **Upstream repo:** https://github.com/ViNet-Saliency/vinet_v2
- **License:** **CC BY-NC-SA 4.0** — https://creativecommons.org/licenses/by-nc-sa/4.0/

**Attribution (required by CC BY-NC-SA 4.0):**

> ViNet-S / ViNet++ saliency weights (c) 2025 Rohit Girmaji, Siddharth Jain, Bhav Beri,
> Sarthak Bansal, Vineet Gandhi (IIIT Hyderabad). "Minimalistic Video Saliency Prediction
> via Efficient Decoder & Spatio-Temporal Action Cues", ICASSP 2025 (arXiv:2502.00397).
> Licensed under CC BY-NC-SA 4.0. Redistributed here, re-containered to safetensors with
> tensor values unchanged, under the same CC BY-NC-SA 4.0 license (ShareAlike) for
> non-commercial use.

The DHF1K visual-only checkpoint is the general saliency model (no audio / no face
dependency), appropriate for subject-tracking crops.

---

## `transnetv2.safetensors` — TransNetV2 shot/scene-boundary detector

- **File size:** 30,481,608 bytes
- **sha256:** `e2877ef6750ccbb3f02256bb4b5f4f53035111677be641d56b9723af499f881d`
- **Tensors:** 90, exact tensor-equality vs original VERIFIED
- **Original source (bytes obtained from):** HuggingFace mirror
  https://huggingface.co/Sn4kehead/TransNetV2 — `transnetv2-pytorch-weights.pth`
  (30,508,183 bytes, sha256 `834b10f25ae9e1b4e4f2652fe2843bd2b1388057a435d68b7c52635578fcc04d`).
- **Upstream (canonical) repo:** https://github.com/soCzech/TransNetV2 (**MIT**). The
  PyTorch weights are a derived artifact of the upstream TensorFlow SavedModel via the
  repo's `convert_weights.py`.
- **License:** **MIT** (soCzech/TransNetV2). Note: the Sn4kehead mirror card labels its
  copy `apache-2.0`; both MIT and Apache-2.0 permit redistribution with attribution. The
  authoritative upstream license for these weights is MIT.

**Attribution:**

> TransNet V2 (c) Tomas Soucek & Jakub Lokoc. "TransNet V2: An Effective Deep Network
> Architecture for Fast Shot Transition Detection." Source:
> https://github.com/soCzech/TransNetV2 (MIT License). PyTorch weights converted from the
> upstream TensorFlow SavedModel; re-containered here to safetensors with tensor values
> unchanged.

---

## Verification recipe (reproducible)

```python
from safetensors.torch import load_file
import torch
sd = load_file("vinet-s-saliency.safetensors")   # or transnetv2.safetensors
# sd is a flat {str: torch.Tensor}; load_state_dict directly. No torch.load / pickle.
```