README.md

---
license: cc-by-nc-sa-4.0
tags:
  - third-eye
  - image-restoration
  - video-restoration
  - super-resolution
  - denoise
  - deblur
  - colorization
  - frame-interpolation
library_name: pytorch
---

# Third Eye - Model Weights Bundle

Reliable mirror of AI model weights used by [Third Eye](https://github.com/Jacid23/Third_Eye), a media organizer with hidden editing dimensions.

These weights are downloaded automatically by `scripts/fetch_models.py` during installation. Self-hosting them here removes dependency on upstream Google Drive links and unreliable community mirrors.

## License

The bundle is tagged **CC BY-NC-SA 4.0** — the most restrictive license among the included models. By using these weights you agree to:

- **Non-commercial use only**
- Provide **attribution** to the original authors (listed below)
- Distribute any derivatives under the **same license**

## Files and Attribution

Every weight in this repo is a verbatim copy of the file released by its original author. Original sources and licenses below.

### Denoise / Deblur (NAFNet)

- `NAFNet-SIDD-width64.pth` — denoise model (SIDD dataset)
- `NAFNet-REDS-width64.pth` — deblur model (REDS dataset)
- `NAFNet-GoPro-width64.pth` — deblur model (GoPro dataset, alternative to REDS)

**Authors:** Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, Jian Sun (Megvii Research)
**Upstream:** https://github.com/megvii-research/NAFNet
**License:** MIT
**Paper:** "Simple Baselines for Image Restoration" (ECCV 2022)

### Frame Interpolation (RIFE)

- `flownet.pkl` — RIFE 4.6 weights

**Authors:** Zhewei Huang et al. (Practical-RIFE team)
**Upstream:** https://github.com/hzwer/Practical-RIFE
**License:** MIT (code) / non-commercial (weights, per author note)
**Paper:** "Real-Time Intermediate Flow Estimation for Video Frame Interpolation"

### Community RRDBNet Upscale Models

**4x variants:**
- `4x-UltraSharp.pth` — community upscale model by Kim2091
- `foolhardy_Remacri.pth` — community model by foolhardy
- `RealisticRescaler_100000_G.pth` — community upscale model
- `4x-UniScale-Balanced-72000g.pth` — UniScale community variant
- `4x-UniScale-Strong-42400g.pth` — UniScale community variant
- `4xJaypeg90.pth` — JPEG-focused 4x cleanup upscaler
- `4xLSDIRplus.pth` — LSDIR dataset upscaler
- `4xLSDIRplusR.pth` — LSDIR refined variant
- `CountryRoads_377000_G.pth` — general-purpose community upscaler
- `NMKD-Superscale-SP_178000_G.pth` — NMKD standard print
- `NMKDSuperscale_Artisoft_120000_G.pth` — NMKD artistic-soft
- `A_ESRGAN_Single.pth` — A-ESRGAN single-pass
- `Filmify4K_v2_325000_G.pth` — film-look upscaler

**8x variants:**
- `8x_NMKD-Superscale_150000_G.pth` — NMKD general 8x
- `8x_NMKD-Typescale_175k.pth` — NMKD optimised for text/UI
- `TGHQFace8x_500k.pth` — face-specific 8x

**1x detail enhancers:**
- `x1_ITF_SkinDiffDetail_Lite_v1.pth` — skin texture enhancement

**Upstream catalog:** https://openmodeldb.info/
**License:** CC BY-NC-SA 4.0 (community convention for ESRGAN-derived models)

Architecture is RRDBNet from Real-ESRGAN. Original Real-ESRGAN architecture:
- **Authors:** Xintao Wang et al. (Tencent ARC Lab)
- **Upstream:** https://github.com/xinntao/Real-ESRGAN
- **License:** BSD-3-Clause

### SwinIR (Swin Transformer Image Restoration)

Initial set wired through the engine:

- `classicalSR_DF2K_s64w8_SwinIR-M_x4.pth` — classical 4x super-resolution
- `classicalSR_DF2K_s64w8_SwinIR-M_x2.pth` — classical 2x super-resolution
- `lightweightSR_DIV2K_s64w8_SwinIR-S_x4.pth` — lightweight 4x (smaller/faster)
- `realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth` — real-world 4x (BSRGAN-trained GAN)
- `colorCAR_DFWB_s126w7_SwinIR-M_jpeg40.pth` — JPEG artifact removal (qf≈40)
- `colorDN_DFWB_s128w8_SwinIR-M_noise25.pth` — color denoise (sigma=25)

**Authors:** Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, Radu Timofte
**Upstream:** https://github.com/JingyunLiang/SwinIR
**License:** Apache 2.0
**Paper:** "SwinIR: Image Restoration Using Swin Transformer" (ICCVW 2021)

Additional SwinIR checkpoints (JPEG qf=10/20/30, noise=15/50, grayscale variants, x3, x8) are available from the upstream releases and can be wired with one MODEL_CONFIGS entry each — the architecture supports all of them.

### Transformer Upscale Models (DAT / HAT-L / DRCT-L)

- `4xFFHQDAT.pth` — DAT architecture, trained on FFHQ
- `4xFaceUpSharpDAT.pth` — DAT, face sharpener
- `4xLSDIRDAT.pth` — DAT, LSDIR dataset
- `4xNomos8kHAT-L_otf.pth` — HAT-L architecture
- `4xNomos2_hq_drct-l.pth` — DRCT-L architecture

**Upstream catalog:** https://openmodeldb.info/
**License:** CC BY-NC-SA 4.0 (community convention)

These are mirrored here for download convenience, but Third Eye's engine does not yet implement the DAT, HAT-L, or DRCT-L architectures. They will be wired up in a future engine update.

Original transformer architecture papers:
- **DAT:** "Dual Aggregation Transformer for Image Super-Resolution" (ICCV 2023)
- **HAT:** "Activating More Pixels in Image Super-Resolution Transformer" (CVPR 2023)
- **DRCT:** "DRCT: Saving Image Super-Resolution away from Information Bottleneck"

## Usage

Download programmatically via the Third Eye installer:

```bat
install.bat
```

Or directly:

```bash
wget https://huggingface.co/Jacid23/third-eye-models/resolve/main/NAFNet-SIDD-width64.pth
```

## Source Code

Third Eye source: https://github.com/Jacid23/Third_Eye

Model download script: `scripts/fetch_models.py`

## Acknowledgements

All credit for the models goes to their original authors and research teams. This repository exists only to provide reliable download mirrors for an open-source application that integrates these models. No modifications have been made to any weight file.