open-photo-ai

Open Photo AI Models

This repository contains a collection of AI models used by Open Photo AI for photo enhancement tasks, including face detection, face recovery, image upscaling, denoising, sharpening, light adjustment, and color balance. Each model is designed to perform a specific function to improve the quality of photographs.

Models

Denoise

• Stockholm: based on NAFNet (SIDD-width64) with fixed shape of 256x256, FP32 and FP16 precisions.
• Gothenburg: based on Restormer (Real Denoise) with fixed shape of 256x256, FP32 and FP16 precisions.
• Malmö: based on Restormer (Derain) with fixed shape of 256x256, FP32 and FP16 precisions.

Detection

• New York: based on RetinaFace (ResNet34), with fixed shape of 640x640, FP32 precision.
• Los Angeles: based on Segment Anything, with dynamic shape, INT8 precision.

Face Recovery

• Athens: based on CodeFormer, with fixed shape of 512x512, FP32 and FP16 precisions.
• Santorini: based on GFPGAN, with fixed shape of 512x512, FP32 and FP16 precisions.

Colorization

• Delhi: based on DDColor (Natural) with fixed shape of 512x512, FP32 and FP16 precisions.
• Mumbai: based on DDColor (Artistic) with fixed shape of 512x512, FP32 and FP16 precisions.
• Jaipur: based on DeOldify (Stable) with fixed shape of 560x560, FP32 and FP16 precisions.

Light Adjustment

• Paris: based on IAT with fixed shape of 1024x1024, FP32 and FP16 precisions.
• Lyon: based on CIT-EC with fixed shape of 1024x1024, FP32 and FP16 precisions.

Color Balance

• Rio: based on Deep White-Balance with fixed shape of 656×656, FP32 and FP16 precisions.
• São Paulo: based on Mixed-Illuminant WB with fixed shape of 656×656, FP32 and FP16 precisions.

Sharpen

• Moscow: based on Restormer (Focus Deblur) with fixed shape of 256x256, FP32 and FP16 precisions.
• St. Petersburg: based on NAFNet (GoPro-width64) with fixed shape of 256x256, FP32 and FP16 precisions.
• Novgorod: based on Restormer (Motion Deblur) with fixed shape of 256x256, FP32 and FP16 precisions.

Upscale

• Tokyo: based on SwinIR, with fixed shape of 256x256, FP32 and FP16 precisions.
• Kyoto: based on Real-ESRGAN (General), with fixed shape of 256x256, FP32 and FP16 precisions.
• Saitama: based on Real-ESRGAN (Anime), with fixed shape of 256x256, FP32 and FP16 precisions.
• Osaka: based on SeedVR2, with dynamic VAE and fixed shape DiT of 960x960, FP16 and INT8 precisions.

Object Removal

• Dubai: coming soon...