MinImagen

A Minimal implementation of the Imagen text-to-image model.

For a tutorial on building this model, see here.

Given a caption of an image, Imagen will generate an image that reflects the caption. The model is a simple cascading diffusion model, using a T5 text encoder to encode the captions which conditions a base image generator, and then a sequence of super-resolution models.

In particular, two notable contributions are the developments of:

1. Noise Conditioning Augmentation, which noises low-resolution conditioning images in the super-resolution models, and
2. Dynamic Thresholding which helps prevent image saturation at high classifier-free guidance weights.

See How Imagen Actually Works for a detailed explanation of Imagen's operating principles.

Attribution Note

This implementation is largely based on Phil Wang's Imagen implementation.

Installation

$ pip install minimagen

Documentation

Documentation can be found here

Usage

A minimal usage:

import torch
from minimagen.Imagen import Imagen
from minimagen.Unet import Unet, Base, Super
from minimagen.t5 import t5_encode_text, get_encoded_dim
from torch import optim

# Name of the T5 encoder to use
encoder_name = 't5_small'

# Text captions of training images
train_texts = [
    'a pepperoni pizza',
    'a man riding a horse',
    'a Beluga whale',
    'a woman rock climbing'
]

# Training images (side length equal to Imagen final output image size)
train_images = torch.randn(4, 3, 64, 64)

# Create the Imagen instance
enc_dim = get_encoded_dim(encoder_name)
unets = (Base(text_embed_dim=enc_dim), Super(text_embed_dim=enc_dim))
imagen = Imagen(unets=unets, image_sizes=(32, 64), timesteps=10)

# Create an optimzier
optimizer = optim.Adam(imagen.parameters())

# Train the U-Nets in Imagen
for j in range(10):
    for i in range(len(unets)):
        optimizer.zero_grad()
        loss = imagen(train_images, texts=train_texts, unet_number=i)
        loss.backward()
        optimizer.step()

# Sample captions to generate images for
sample_captions = [
    'a happy dog',
    'a big red house',
    'a woman standing on a beach',
    'a man on a bike'
]

# Generate images
images = imagen.sample(texts=sample_captions, cond_scale=3., return_pil_images=True)

# Save images
for idx, img in enumerate(images):
    img.save(f'Generated_Image_{idx}.png')

Text embeddings and masks can be precomputed, and Unets parameters can be specified rather than using Base and Super:

train_encs, train_mask = t5_encode_text(train_texts, name=encoder_name)

enc_dim = get_encoded_dim(encoder_name)

base_unet = Unet(
    dim=32,
    text_embed_dim=enc_dim,
    cond_dim=64,
    dim_mults=(1, 2, 4),
    num_resnet_blocks=2,
    layer_attns=(False, False, True),
    layer_cross_attns=(False, False, True),
    attend_at_middle=True
)

super_res_unet = Unet(
    dim=32,
    text_embed_dim=enc_dim,
    cond_dim=512,
    dim_mults=(1, 2, 4),
    num_resnet_blocks=(2, 4, 8),
    layer_attns=(False, False, True),
    layer_cross_attns=(False, False, True),
    attend_at_middle=False
)

# Create Imagen instance
imagen = Imagen((base_unet, super_res_unet), image_sizes=(32, 64), timesteps=10)