README.md

metadata

title: AlignDRAW
emoji: 🖍️
colorFrom: gray
colorTo: yellow
sdk: gradio
sdk_version: 6.16.0
python_version: '3.10'
app_file: app.py
hardware: zero-a10g
pinned: false
short_description: One of the first text-to-image models (2015), resurrected

🖍️ AlignDRAW

A resurrection of AlignDRAW (Mansimov, Parisotto, Ba, Salakhutdinov — Generating Images from Captions with Attention, ICLR 2016) — one of the very first neural text-to-image models, predating GAN and diffusion text-to-image.

Type a caption and watch a recurrent-attention VAE draw it onto a 32×32 canvas over 32 steps, attending to the caption words at each step. The results are tiny and blurry — but their colour and layout track the caption (red for "stop sign", yellow for "school bus", green fields for "elephants") — exactly the fidelity this model had in 2015.

The original Theano code shipped no pretrained weights, so this is a faithful PyTorch re-implementation trained from scratch on the original COCO 32×32 caption data (still online at cs.toronto.edu/~emansim). Runs on ZeroGPU. Weights + model + dictionary hosted at mediasynthesismuseum/aligndraw.

Credit: Elman Mansimov, Emilio Parisotto, Jimmy Ba, Ruslan Salakhutdinov; built on DRAW (Gregor et al., 2015).