AI-Powered Super Resolution · CS360 Artificial Intelligence · Semester 4
Real-ESRGAN
4× Upscale
Artwork Optimized
Deep Learning
CS360
Rajiv Gandhi Institute of Petroleum Technology Jais, Amethi
""")
with gr.Tabs(elem_classes=["tab-nav"]):
# ── TAB 1: ENHANCE ───────────────────────────────────────────────────
with gr.Tab("✦ Enhance"):
gr.HTML("""
⚡ For fast demo results — use small artwork images (under 300×300px).
Try anime thumbnails, pixel art, game sprites, or low-res manga panels.
Processing time on CPU: ~8–15s for small images, ~30–60s for large ones. Good test sources:
Google Images → Tools → Size → SmallPinterest low-res artGame sprite sheets
· Anime, manga & illustrations → best output
· Pixel art → use 4× for crisp upscale
· Photos work too but are slower
· Start with 4× — it is the supported upscale
· The queue timer is normal — it's processing!
Good test images
anime thumbnailpixel art spritemanga panelgame screenshotlow-res painting
""")
with gr.Column(scale=1, min_width=320):
gr.HTML("""
⏳ Processing takes 10–60s on CPU · The queue timer below the image is normal · Please wait
The model analysed every 256×256 tile of your image,
ran it through 6 RRDB neural network blocks,
and reconstructed fine detail at up to 4× the original resolution —
synthesizing sharp edges and textures that weren't in the original.
""")
enhance_btn.click(fn=enhance, inputs=[input_image, scale], outputs=[output_image, info_box])
clear_btn.click(fn=lambda: (None, None, ""), outputs=[input_image, output_image, info_box])
# ── TAB 2: THE MODEL ─────────────────────────────────────────────────
with gr.Tab("◈ The Model"):
gr.HTML(f"""
Real-ESRGAN
Real-Enhanced Super-Resolution Generative Adversarial Network —
a deep learning model that doesn't just stretch pixels,
it invents the fine detail that was lost in compression or downsizing.
Think of it as an AI that has studied millions of high-resolution images
and learned exactly what sharp edges and crisp textures look like —
then applies that knowledge to reconstruct your blurry image from scratch.
4×
Max Upscale
6B
RRDB Blocks
17M
Parameters
3
Loss Functions
What is a GAN? — The Forger & The Detective
{gan_diagram}
🎨
Generator — The Art Forger
Takes a blurry, low-res input and tries to produce a crisp,
photorealistic high-res version. It starts terrible and gets better
every time the discriminator catches it out. Its job is to make fakes
so good that even an expert can't tell them apart from the real thing.
🔍
Discriminator — The Detective
A second network trained purely to spot fakes. It sees both real
high-res images and the generator's outputs, and tries to label each one.
Its feedback is what drives the generator to keep improving — without it,
the generator would just produce blurry averages.
🔄
Adversarial Training Loop
The two networks compete millions of times. The forger gets better at
fooling, the detective gets better at catching. After enough rounds,
the generator has learned to produce images so convincing the
discriminator genuinely can't tell what's real.
✨
Why Not Just Bicubic?
Bicubic interpolation averages nearby pixels — it makes mathematical sense
but looks blurry. GANs instead ask "what would a real high-res image look like
here?" and synthesise convincing detail. The difference is
interpolation vs learned perception.
ESRGAN Architecture & Pipeline
{esrgan_diagram}
🧱
RRDB — Residual-in-Residual Dense Blocks
The core building unit. Each RRDB contains dense connections where every layer
receives input from all previous layers — letting gradients flow freely
during training and allowing the network to learn extremely fine texture patterns.
Stacking 6 of these creates a deep feature extractor without gradient degradation.
🔲
Tile-Based Inference
Images are split into overlapping 256×256 tiles processed individually,
then stitched back. The overlap (tile padding) prevents visible seams.
This lets the model run on consumer hardware — no GPU needed —
and makes RAM usage predictable regardless of image size.
📐
Pixel Shuffle Upscaling
Rather than resizing first (causing blur), the network works at the original
resolution throughout. At the very end, a pixel-shuffle layer rearranges
learned sub-pixel channels into the final upscaled output —
preserving all the sharpness built up through the RRDB blocks.
🎌
Anime 6B Variant
Fine-tuned on anime and illustrated art using only 6 RRDB blocks
(vs 23 in the full model) for faster CPU inference.
It's optimised for clean linework, flat colour shading,
and the vibrant palette of illustrated content — exactly what we use here.
Loss Functions — How the Model Learns
{loss_diagram}
📏
Pixel Loss (L1 / MSE)
Compares the output image to the real high-res image pixel by pixel.
Simple but essential — it ensures the overall structure and colour
are correct. On its own it produces blurry results because it
averages all possible sharp outputs into one smooth prediction.
⚔️
Adversarial Loss
Measures how well the generator fools the discriminator.
This is the engine of sharpness — it pushes the generator to produce
outputs that look statistically indistinguishable from real high-res images,
forcing it to synthesise convincing textures and edges.
🧠
Perceptual Loss (VGG Features)
Uses a pre-trained VGG network to compare images at a feature level
rather than pixel level. Instead of asking "are these pixels identical?",
it asks "do these images contain the same objects and structures?"
This preserves semantic meaning even when pixel values differ.
⚖️
Combined Loss = Best of All Three
The final loss is a weighted sum: pixel loss keeps structure correct,
adversarial loss adds sharpness and realism, perceptual loss preserves
semantic content. Together they prevent the three failure modes:
structural wrong output, blurriness, and semantic distortion.
The Key Insight
Traditional methods ask "what pixel value makes mathematical sense here?"
Real-ESRGAN asks "what would a human photographer actually have captured here?"
That shift — from interpolation to learned perception — is why the output looks
sharp and natural rather than smooth and artificial.
The three-loss training framework ensures the model is simultaneously accurate,
sharp, and semantically meaningful.
""")
# ── TAB 3: TEAM ─────────────────────────────────────────────────────
with gr.Tab("◉ Our Team"):
gr.HTML("""
CourseCS360 · Artificial Intelligence
Semester4th Semester · 2025–26
InstructorDr. Susham Biswas
InstituteRGIPT, Jais, Amethi
⭐ Team Leader
Anurag Sharma
24MC3008
Project architecture, system integration, environment setup, and final submission. Coordinates all modules into a cohesive working application.
Model Research
Nityansh Pant
24MC3033
Deep-dives into Real-ESRGAN architecture, RRDB block design, and GAN theory. Responsible for literature review and model selection rationale.
UI / UX Developer
Vaibhav
24MC3059
Designs and implements the Gradio interface — layout, custom CSS theming, tab structure, and overall visual experience of the application.
Image Preprocessing
Shubhayu Brahmachari
24MC3046
Handles input validation, format conversion, colour space management, and post-processing of enhanced outputs before display.
Inference Engine
Himanshu Sachdeva
24MC3021
Implements model loading, tiling strategy, upsampler configuration, and CPU inference pipeline for efficient processing.
Testing & Evaluation
Arjit Anand
24MC3009
Conducts before/after quality comparisons, documents metrics, and validates model performance across different artwork styles.
Documentation
Taksh Agarwal
24MC3051
Writes the README, inline code documentation, report structure, and technical write-up covering methodology and results.
Demo & Presentation
Utkarsh Dixit
24MC3057
Records the video demonstration, coordinates individual narration segments, and prepares the final presentation for submission.
