🖼️ Thanawanit's Upscale ONNX Models Collection

This repository contains a curated set of 8 ONNX models optimized for image and video upscaling. The collection includes models tuned for anime line-art, live-action content, denoising, sharpening, and compression artifact removal.

📦 Included Models

Model File	Scale	Category / Type	Original License	Notes
Adore_2x.onnx	2x	Anime / Line Art	Unknown	Fast compact model; clears blur and sharpens anime line art.
Ani4Kv2_2x.onnx	2x	Ani4Kv2 Project	CC BY-NC 4.0	Compact variant; excellent for preserving depth-of-field and smooth textures.
Balanced_2x.onnx	2x	General Purpose	Unknown	Fast, versatile, and balanced general-purpose upscaler.
LiveActionV1_2x.onnx	2x	Live-Action / SPAN	Unknown	SPAN architecture model tuned for live-action videos, real faces, and movies.
RealESRGAN_2x.onnx	2x	Xintao Wang	BSD 3-Clause	Compact Real-ESRGAN; fast, stable, and reliable.
Strong_v1.5.onnx	2x	Rescue / De-artifact	Unknown	"Rescue" model for heavily compressed or low-quality inputs.
Strong_v2.5.onnx	2x	Smooth / VFX	Unknown	Smooth sharpness; great for VFX-heavy edits and video.
Strong_v3.onnx	2x	Sharp / Detail	Unknown	Maximum edge sharpness and detail enhancement.

Important: Several models (Adore, Balanced, LiveAction, Strong series) have unknown original licenses. They are included here for research and personal use only. If you are the original creator and would like them removed or properly attributed, please open an issue.

📜 Overall License

This collection as a whole is distributed under the Creative Commons Attribution‑NonCommercial‑ShareAlike 4.0 International license (CC BY‑NC‑SA 4.0).

However, individual models retain their original licenses (see table above). Commercial use may require separate permission from the original authors.

🚀 Usage

All models are provided in the ONNX format and can be used directly with ONNX Runtime.

Python Example

import onnxruntime as ort
import cv2
import numpy as np

# Load model (e.g., Balanced_2x.onnx)
sess = ort.InferenceSession("Balanced_2x.onnx", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])

# Preprocess image
img = cv2.imread("input.jpg").astype(np.float32) / 255.0
img = img.transpose(2, 0, 1)[np.newaxis, ...]

# Run inference
input_name = sess.get_inputs()[0].name
output_name = sess.get_outputs()[0].name
output = sess.run([output_name], {input_name: img})[0]

# Postprocess image
output = (output.squeeze().transpose(1, 2, 0) * 255).clip(0, 255).astype(np.uint8)
cv2.imwrite("output.png", output)