ArIES, IIT Roorkee

Text-to-Image GAN

A stacked GAN generating photorealistic birds from text descriptions

PythonPyTorchGANs

Problem: Generating a photorealistic image from a sentence is a resolution problem as much as a semantic one. A single generator asked to go straight from text to a sharp image tends to produce something that is either blurry or ignores half the description.
My role: Sole implementer, as part of ArIES at IIT Roorkee.
Approach: Stacked the generation in stages rather than attempting it in one — an early stage fixes shape and colour from the text embedding, a later stage refines detail conditioned on both the text and the previous output. More moving parts to train and tune, but each stage has a tractable job.
Outcome: Trained on CUB-200-2011 — 200 bird species, 11,788 images — producing recognisable, text-faithful samples.

Why staged generation

The failure mode of single-stage text-to-image at this scale, and what stacking buys you.

How the text embedding enters each stage, and what changes between them.

Sample outputs, and where it falls down. The limitations section is what makes this read as engineering rather than a course submission.