Upscale and Refine Models
A curated collection of ONNX models for image upscaling, denoising, deblurring, colorization, segmentation, and other image restoration tasks. All models are in ONNX format for easy cross-platform inference.
Models
Upscaling
| Model |
Scale |
Description |
4xNomosWebPhoto_atd |
4x |
High-quality photo upscaler |
SwinIR-4x-GAN |
4x |
SwinIR GAN-based upscaler |
SPAN-4x |
4x |
Lightweight efficient upscaler |
realcugan-2x-* |
2x |
Real-CUGAN variants (conservative, no-denoise, denoise 1x/2x/3x) |
realcugan-3x-* |
3x |
Real-CUGAN variants (conservative, denoise 3x) |
realcugan-4x-* |
4x |
Real-CUGAN variants (conservative, no-denoise, denoise 3x) |
realesrgan-general-* |
4x |
Real-ESRGAN for general photos (fast / plus) |
realesrgan-anime-* |
4x |
Real-ESRGAN for anime images (fast / plus) |
Denoising & Artifact Removal
| Model |
Description |
1xDeNoise_realplksr_otf |
General denoising |
1xDeJPG_realplksr_otf |
JPEG artifact removal |
1xDeH264_realplksr |
H.264 compression artifact removal |
SCUNet-PSNR |
Blind denoising (PSNR-oriented) |
SCUNet-GAN |
Blind denoising (GAN-oriented, perceptual) |
dncnn-color-blind |
Blind color image denoising |
NAFNet-SIDD-width64 |
Real-world denoising |
Deblurring
| Model |
Description |
1x-hurrdeblur-superultracompact |
Lightweight deblurring |
NAFNet-GoPro-width64 |
Motion deblurring (GoPro) |
NAFNet-REDS-width64 |
Motion deblurring (REDS) |
maxim-deblurring |
MAXIM deblurring |
maxim-deblurring-reds |
MAXIM deblurring (REDS) |
maxim-deblurring-realblur-j |
MAXIM deblurring (RealBlur-J) |
maxim-deblurring-realblur-r |
MAXIM deblurring (RealBlur-R) |
Image Enhancement & Restoration
| Model |
Description |
maxim-enhancement |
Low-light enhancement |
maxim-retouching |
Photo retouching |
maxim-dehazing-indoor |
Indoor dehazing |
maxim-dehazing-outdoor |
Outdoor dehazing |
maxim-deraining |
Deraining |
maxim-deraining-raindrop |
Raindrop removal |
maxim-denoising |
MAXIM denoising |
Refinement
| Model |
Description |
1x-aniscale2-refiner |
Anime refinement |
1x-artclarity |
Art clarity enhancement |
1x-ghibli-grain |
Ghibli-style film grain |
Colorization
| Model |
Description |
ddcolor-tiny |
DDColor colorization (lightweight) |
ddcolor-large |
DDColor colorization (high quality) |
deoldify / deoldify_fp16 |
DeOldify colorization (fp32 / fp16) |
deoldify_fixed / deoldify_fp16_fixed |
DeOldify with fixes (fp32 / fp16) |
Segmentation & Background Removal
| Model |
Description |
RMBG-1.4 |
Background removal |
BiRefNet-lite-fp16 |
Bilateral reference segmentation (fp16) |
mediapipe-selfie-segmenter |
Person segmentation |
mediapipe-multiclass-segmenter |
Multi-class segmentation |
Classification
| Model |
Description |
anime_real_cls-mobilenetv3 |
Anime vs. real photo classifier |
nima-mobilenet-quality |
Image quality assessment (NIMA) |
Usage
All models are in ONNX format and can be loaded with any ONNX-compatible runtime:
import onnxruntime as ort
session = ort.InferenceSession("realesrgan-general-plus.onnx")
result = session.run(None, {"input": input_tensor})
License
This repository contains models from multiple open-source projects, each with its own license. See LICENSE for details.
