File size: 2,566 Bytes
04bddd6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | ---
license: bsd-3-clause
library_name: libreyolo
tags:
- super-resolution
- image-to-image
- quicksrnet
- real-time
pipeline_tag: image-to-image
---
# LibreQuickSRNetm2-restore
QuickSRNet Medium 2x compact real-time super-resolution, packaged for
LibreYOLO. It has 50,604 parameters and produces an RGB image at twice the
input height and width.
## Source
The architecture and official checkpoint come from
[quic/aimet-model-zoo](https://github.com/quic/aimet-model-zoo/tree/1bd2bf5b17cdda9251437c444009b29e1a25054b/aimet_zoo_torch/quicksrnet)
at commit `1bd2bf5b17cdda9251437c444009b29e1a25054b`, BSD-3-Clause.
Copyright (c) 2022 Qualcomm Innovation Center, Inc.
The maintained Qualcomm integration was verified against
[qualcomm/ai-hub-models](https://github.com/qualcomm/ai-hub-models/tree/16dbeb5e2805d4ada7218026de72e36878717d46/src/qai_hub_models/models/quicksrnetmedium)
at commit `16dbeb5e2805d4ada7218026de72e36878717d46`, BSD-3-Clause.
Official source artifact:
`quicksrnet_medium_2x_checkpoint_float32.pth.tar`
SHA-256:
`a0d176b40a649e45a176c3b53f45e0237015f4f2c17b157ef5c81e38c4442a0d`
The model card records DIV2K as the training dataset.
## Modifications
The 14 learned state-dict tensors are unchanged. Conversion discards the
training-only epoch, optimizer, PSNR, and SSIM objects, then wraps the tensors
in the LibreYOLO v1.0 checkpoint schema with `task=restore`, `size=m2`, and
`scale=2`. FP32 tensor output matches the pinned upstream implementation
exactly (`max_abs_diff == 0`).
Converted checkpoint SHA-256:
`3f779b461d200704ab82904a110aaa37f41e8258b285e02b7e413b955deb150b`
## Usage
```python
from libreyolo import LibreYOLO
model = LibreYOLO("LibreQuickSRNetm2-restore.pt")
result = model.predict("small.jpg")
print(result.restore_scale) # 2
result.save("upscaled.png")
```
Native PyTorch prediction accepts arbitrary positive image dimensions. Dynamic
spatial ONNX and fixed-canvas TorchScript export are supported.
## Reference latency
Synchronized model-forward timings on an NVIDIA GeForce RTX 5070 Ti with
PyTorch 2.11.0 and CUDA 12.8, batch 1, cuDNN benchmarking enabled, 10 warmups,
and 30 timed iterations:
| Input to output | Precision | Median | p95 |
|---|---|---:|---:|
| 360p to 720p | FP32 | 1.746 ms | 1.770 ms |
| 360p to 720p | FP16 | 0.937 ms | 0.966 ms |
| 720p to 1440p | FP32 | 10.536 ms | 10.931 ms |
| 720p to 1440p | FP16 | 5.437 ms | 5.855 ms |
Image decoding, transfer, and result conversion are excluded.
## License
BSD 3-Clause. See [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE).
|