Legitking4pf/Real-ESRGAN-Upscaler

Model Card for Model ID

Legit-UpscaleAI is a GAN-based image super-resolution and enhancement model optimized for real-world images such as mobile photos, compressed web images, and user-generated content. The model focuses on perceptual quality, edge sharpness, and texture recovery rather than synthetic over-sharpening.

Model Details

Model Description

It is designed to be deployable across environments (mobile, desktop, server) and supports device-aware inference strategies.

• Developed by: King Marvis
• Funded by [optional]: Independent / Self-funded
• Shared by [optional]: King Marvis
• Model type: Image Super-Resolution & Image Enhancement
• Language(s) (NLP): Not applicable (vision-only model)
• License: MIT
• Finetuned from model [optional]: Real-ESRGAN family

Model Sources [optional]

• Repository: https://huggingface.co/Legitking4pf/Real-ESRGAN-Upscaler
• Paper [optional]: Not published
• Demo [optional]: Planned (Web & Mobile)

Uses

Direct Use

This model can be used directly to:

• Upscale low- and mid-resolution images to higher resolutions
• Enhance clarity and perceptual quality of compressed images
• Restore noisy or degraded images for display or printing

Typical users include developers, content creators, and researchers working with image enhancement pipelines.

Downstream Use [optional]

Downstream use cases include:

• Mobile photo enhancement apps
• Desktop image editing software
• Video frame upscaling pipelines
• AI-powered creative tools
• Cloud-based image processing APIs

Out-of-Scope Use

This model is not suitable for:

• Facial recognition or biometric identification
• Medical imaging or diagnostics
• Forensic reconstruction
• Surveillance or identity inference

Bias, Risks, and Limitations

As a super-resolution model, Legit-UpscaleAI may hallucinate fine textures in some cases, particularly when scaling extremely low-quality inputs. Enhanced outputs should not be treated as ground truth representations of original scenes.

The model may perform inconsistently on non-photographic or heavily stylized images.

Recommendations

Users should avoid using enhanced outputs for legal, medical, forensic, or evidentiary purposes. Conservative scaling factors are recommended for sensitive content.

How to Get Started with the Model

import torch
from PIL import Image
from realesrgan import RealESRGAN

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = RealESRGAN(device, scale=4)
model.load_weights("Legit-UpscaleAI.pth")

input_image = Image.open("input.jpg").convert("RGB")

with torch.no_grad():
    output_image = model.predict(input_image)

output_image.save("output.png")

Training Details

Training Data

Training was conducted using a curated mix of high-resolution and real-world datasets, including DF2K (DIV2K + Flickr2K), LAION high-resolution subsets, Pexels, Shutterstock samples, and Getty Images sample datasets.

Synthetic degradations such as blur, noise, and compression were applied to simulate real-world image conditions.

Training Procedure

Preprocessing [optional]

• Bicubic downsampling
• Gaussian and motion blur
• JPEG compression artifacts
• Noise injection
• Color normalization

Training Hyperparameters

• Training regime: fp16 mixed precision

Speeds, Sizes, Times [optional]

Training was performed over multiple phases using GPU acceleration. Final checkpoints were optimized for inference efficiency.

Evaluation

Testing Data, Factors & Metrics

Testing Data

• DIV2K validation set
• Flickr2K validation
• Real-world mobile photography samples

Factors

• Upscaling factor (×2, ×4, ×8)
• Image content type
• Compression level

Metrics

• PSNR for reconstruction fidelity
• SSIM for structural similarity
• LPIPS for perceptual similarity

Results

Legit-UpscaleAI demonstrates competitive perceptual quality compared to baseline Real-ESRGAN models, particularly on real-world compressed images.

Summary

The model prioritizes visual realism and practical usability over purely synthetic benchmark optimization.

Model Examination [optional]

No formal interpretability analysis has been conducted.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA GPU (RTX / A100 class).
• Hours used: 300–500 GPU hours (estimated).
• Cloud Provider: Mixed local and cloud
• Compute Region: Global
• Carbon Emitted: Not measured

Technical Specifications [optional]

Model Architecture and Objective

Model Architecture and Objective GAN-based super-resolution architecture using Residual-in-Residual Dense Blocks (RRDB), optimized with perceptual and adversarial losses.

Compute Infrastructure

Multi-GPU training with CUDA acceleration.

Hardware

NVIDIA CUDA-compatible GPUs recommended.

Software

• PyTorch
• CUDA
• Hugging Face Hub
• Real-ESRGAN framework

Citation [optional]

BibTeX:

@article{wang2021realesrgan,
  title     = {Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data},
  author    = {Wang, Xintao and Xie, Liangbin and Dong, Chao and Shan, Ying},
  journal   = {Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
  year      = {2021},
  url       = {https://arxiv.org/abs/2107.10833}
}

APA:

King Marvis. (2025). Legit UpscaleAI: Real-World Image Super-Resolution. Hugging Face.

Glossary [optional]

• Super-Resolution: Increasing image resolution using learned mappings.
• GAN: Generative Adversarial Network.
• LPIPS: Learned Perceptual Image Patch Similarity.

[More Information Needed]

More Information [optional]

Future updates will include model variants optimized for mobile and video workflows.

Model Card Authors [optional]

King Marvis

Model Card Contact

Contact via Hugging Face profile