---
license: mit
library_name: pytorch
pipeline_tag: image-to-image
tags:
  - super-resolution
  - onnx
  - pytorch
  - npu
  - game-video
  - minecraft
  - wuthering-waves
---

# LunaSR General MC65-WW35 x2

Tiny x2 luminance super-resolution model specialized for captured game footage. The training patch mixture used approximately 65% Minecraft frames and 35% Wuthering Waves frames.

contains no source videos, extracted frames, faces, or training dataset.

## Model

- Architecture: Dense8-B6
- Parameters: 3,876
- Scale: x2
- Signal: gamma-encoded BT.709 luma
- Output: luma residual, not complete RGB
- PyTorch checkpoint: `pytorch_model.pt`
- Fixed-shape ONNX: `model.onnx`

The ONNX graph accepts `[1, 1, 540, 960]` float32 luma and returns `[1, 1, 1080, 1920]` float32 residual. Compose the final image as:

```text
rgb_hr = clamp(bilinear_x2(rgb_lr) + luma_residual, 0, 1)
```

The same residual is added to R, G, and B to preserve the bilinearly upscaled chroma.

## Quick start

```powershell
python -m pip install -r requirements.txt

python -m scripts.upscale_images `
  --input "input.png" `
  --output "output" `
  --checkpoint pytorch_model.pt `
  --device cuda `
  --save-bilinear
```

Video upscaling requires FFmpeg and FFprobe on PATH:

```powershell
python scripts/upscale_video.py `
  --input "input.mp4" `
  --output "output.mp4" `
  --checkpoint pytorch_model.pt `
  --device cuda `
  --batch-size 16
```

If the source container reports incorrect FPS metadata, normalize the input timestamps first or correct the output timestamps afterward.

## External holdout results

BT.709 luma PSNR improvement over bilinear x2:

| Domain | Baseline | Model | Delta |
|---|---:|---:|---:|
| Minecraft | 40.369 dB | 42.566 dB | +2.197 dB |
| Wuthering Waves | 46.005 dB | 47.874 dB | +1.869 dB |

The 65:35 weighted delta was `+2.082 dB`. See `metrics/model_domain_eval_matrix.json` for the reproducible evaluation record.

## Validation

- ONNX checker: passed
- ONNX Runtime CPU execution: passed
- PyTorch vs ONNX maximum absolute error: `3.8743019104e-7`
- ONNX operators: `Add`, `Conv`, `DepthToSpace`, `Relu`
- Actual Intel NPU/OpenVINO runtime: not tested

RTX 4060 Ti FP32 measurements for the CUDA-resident RGB composition path:

| Input -> output | Batch 1 throughput |
|---|---:|
| 256x256 -> 512x512 | about 1,148 fps |
| 512x512 -> 1024x1024 | about 534 fps |

These are implementation-specific local measurements, not guaranteed device performance.

## Limitations

- Single-frame model; no explicit temporal loss or recurrent state.
- Trained primarily on two game domains and may oversharpen unrelated animation, text, faces, or photographic content.
- The provided ONNX uses a fixed input shape for current NPU compiler compatibility.
- RGB reconstruction preserves bilinear chroma and only predicts shared luma detail.

## Files

- `model.onnx`: fixed 960x540 residual-only ONNX
- `model.json`: deployment contract
- `model.verification.json`: numerical verification
- `pytorch_model.pt`: original training checkpoint
- `lunasr/model.py`: architecture definition
- `scripts/`: image, video, export, and CUDA benchmark utilities
- `SHA256SUMS.txt`: file integrity manifest