Implementation:Junyanz Pytorch CycleGAN and pix2pix ImagePool Query

**Metadata**
Knowledge Sources	pytorch-CycleGAN-and-pix2pix
Domains	Image-to-Image Translation, GAN Training Stabilization, Generative Adversarial Networks
Last Updated	2026-02-09

Overview

The ImagePool class in util/image_pool.py implements a fixed-size buffer of previously generated images used to stabilize discriminator training in CycleGAN.

Description

The ImagePool class maintains an internal list of up to pool_size image tensors. When query() is called with a batch of newly generated images, the pool applies a stochastic replacement policy: each image in the batch either passes through unchanged or is swapped with a randomly selected historical image from the pool. The returned batch is then used to compute the discriminator loss.

If pool_size is set to 0, the pool is disabled and query() simply returns the input images unchanged. This is the behavior used by pix2pix.

Usage

Instantiated in CycleGANModel.__init__ with pool_size=50 (default). Called in CycleGANModel.backward_D_basic to retrieve images for discriminator updates.

Code Reference

Source Location

File	Lines
util/image_pool.py	L5-54

Signature

class ImagePool():
    def __init__(self, pool_size):
        """Initialize the ImagePool class.

        Parameters:
            pool_size (int) -- the size of image buffer; if pool_size=0, no buffer will be created
        """

    def query(self, images):
        """Return an image from the pool.

        Parameters:
            images (torch.Tensor) -- the latest generated images from the generator

        Returns:
            return_images (torch.Tensor) -- images from the buffer (may include historical images)

        By 50/50 chance, the buffer will return a previously stored image
        and insert the current image into the buffer.
        """

Import

from util.image_pool import ImagePool

fake_A_pool = ImagePool(pool_size=50)
fake_B_pool = ImagePool(pool_size=50)

I/O Contract

**Inputs**
Parameter	Type	Description
pool_size	int	Maximum number of images to store in the buffer (0 disables the pool)
images	torch.Tensor	Batch of generated images from the generator, shape (N, C, H, W)

**Outputs**
Output	Type	Description
return_images	torch.Tensor	Batch of images to use for discriminator update; may contain historical images swapped from the pool

Usage Examples

from util.image_pool import ImagePool

# Create pools for CycleGAN (one per direction)
fake_A_pool = ImagePool(pool_size=50)
fake_B_pool = ImagePool(pool_size=50)

# During training, query the pool before discriminator update
fake_B = netG_A(real_A)  # Generate fake B from real A
fake_B_pooled = fake_B_pool.query(fake_B)  # May return historical fake_B

# Use pooled images for discriminator loss
pred_fake = netD_B(fake_B_pooled.detach())
loss_D_B = criterionGAN(pred_fake, False)

# Disabled pool for pix2pix
no_pool = ImagePool(pool_size=0)
same_images = no_pool.query(generated_images)  # Returns input unchanged

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment