DIS

DIS (Direct Image Supersampling) is a lightweight image super-resolution architecture optimized for speed and real-time inference. It has support for PyTorch, ONNX, and TensorRT.

This is the inference and ONNX conversion code. To train a model, you'll want to use traiNNer-redux.

Getting Started

1. Clone the repository:

   git clone https://github.com/Kim2091/DIS
   

2. Install PyTorch with CUDA: Follow the instructions at pytorch.org.

3. Install required packages:

   pip install -r requirements.txt
   

Model Variants

Variant	Parameters (2x)	Description
DIS_Balanced	~269K	Balance of speed and quality
DIS_Fast	~195K	Fastest, recommended

Benchmarks

Configuration: 2x upscale, 720p, FP16 with TensorRT, 2 streams

Model	FPS	PSNR (BHI100)	SSIM (BHI100)	Notes
DIS_Balanced	100	27.44	0.898	Slightly behind Compact, faster
DIS_Fast	137	27.27	0.895	On par with ArtCNN R8F48, 2x faster
ArtCNN R8F48	86	27.25	0.897	Reference model
Compact	78	27.59	0.90	Reference model

Usage

Command-Line Usage

Image upscaling (PyTorch):

python inference.py --input lr.png --output sr.png --scale 4 --fp16

ONNX inference:

python inference.py --input lr.png --output sr.png --model model.onnx --backend onnx

Benchmark:

python inference.py --benchmark --scale 4 --fp16

Tools

Utility scripts are located in the tools/ directory.

Convert PyTorch model to ONNX:

python tools/export_onnx.py --model pretrained_models/model.pth --output model.onnx

• --dynamic: Create a model that supports various input sizes.
• --fp16: Convert the model to FP16 for a speed boost.

TensorRT

The easiest way to use this model with TensorRT is through Vapourkit or VideoJaNai.

Alternatively you can convert to TensorRT manually:

# Dynamic shapes
trtexec --onnx=model_fp16.onnx \
    --minShapes=input:1x3x64x64 \
    --optShapes=input:1x3x256x256 \
    --maxShapes=input:1x3x1024x1024 \
    --saveEngine=model_dynamic.engine \
    --fp16