ONNX models used by AIVideoMagus for OCR, AI upscaling,
face detection, segmentation, restoration, and enhancement workflows.
Optimized for:
CoreML / Apple Neural Engine,
ONNX Runtime,
TensorRT,
and
DirectML.
PP-OCRv5 Mobile Detection
Lightweight OCR text detector optimized for mobile devices.
File: PP-OCRv5_mobile_det_infer.onnx
Best For:
Watermarks, subtitles, UI text, screenshots, signs.
Download
PP-OCRv5 Mobile Recognition
OCR recognition model used after text detection.
File: PP-OCRv5_mobile_rec_infer.onnx
Tip:
Pair with the detection model for full OCR workflows.
Download
RealESRGAN x2
Moderate 2x AI upscaler balancing speed and detail recovery.
File: RealESRGAN_x2.onnx
Download
RealESRGAN x2 FP16
FP16 optimized x2 upscaler for faster inference.
File: RealESRGAN_x2_fp16.onnx
Download
RealESRGAN x2plus FP16
Higher-quality 2x restoration with improved texture synthesis.
File: RealESRGAN_x2plus.fp16.onnx
Download
RealESRGAN x4
Standard 4x AI enhancement model.
File: RealESRGAN_x4.onnx
Download
RealESRGAN x4 FP16
FP16 optimized 4x upscaler for faster hardware acceleration.
File: RealESRGAN_x4_fp16.onnx
Download
RealESRGAN x4plus FP16
Premium 4x restoration model with stronger detail synthesis.
File: RealESRGAN_x4plus.fp16.onnx
Download
RealESRGAN x8
Extreme 8x enhancement model for aggressive enlargement.
File: RealESRGAN_x8.onnx
Download
RealESRGAN x8 FP16
FP16 optimized version of the x8 enhancement model.
File: RealESRGAN_x8_fp16.onnx
Download
Real-ESRGAN x4plus
Original Real-ESRGAN x4plus architecture widely used in restoration pipelines.
File: Real-ESRGAN-x4plus.onnx
Download
YOLOv8 Face
Fast face detection model for masking, tracking, enhancement,
and preprocessing pipelines.
File: yolov8m-face.onnx
Best For:
Real-time face detection and blur targeting.
Download
YOLOv11 Face Models
Newer-generation face detectors with improved robustness,
tracking consistency, and accuracy.
Files:
yolov11n-face.onnx
yolov11s-face.onnx
yolov11m-face.onnx
yolov11l-face.onnx
Variants:
n = Nano / fastest
s = Small
m = Medium
l = Large / highest accuracy
Nano
Small
Medium
Large
RVM MobileNetV3 Models
Real-time video matting models based on Robust Video Matting
with a MobileNetV3 backbone for fast background removal,
alpha extraction, and green-screen style compositing.
Files:
rvm_mobilenetv3_fp32.onnx
rvm_mobilenetv3_fp16.onnx
Variants:
FP32 = Full precision / highest quality
FP16 = Half precision / faster + lower VRAM usage
FP32
FP16
LaMa Inpainting Models
Advanced image inpainting models designed for object removal,
watermark cleanup, hole filling, and scene reconstruction using
Fourier convolution techniques for strong large-area consistency.
Files:
lama_fp32.onnx
lama_fp16.onnx
Variants:
FP32 = Full precision / best reconstruction quality
FP16 = Half precision / faster inference + reduced memory usage
FP32
FP16