Implementation:Junyanz Pytorch CycleGAN and pix2pix Save Images

**Metadata**
Knowledge Sources	pytorch-CycleGAN-and-pix2pix
Domains	Image-to-Image Translation, Visualization, Tensor Processing
Last Updated	2026-02-09

Overview

The save_images function in util/visualizer.py converts model output tensors to image files and adds them as rows in an HTML gallery. It relies on tensor2im and save_image from util/util.py for tensor conversion and file writing.

Description

The save_images function receives a dictionary of visual outputs from the model (keyed by name, e.g., real_A, fake_B), the HTML webpage object, the image path, and display parameters. For each visual:

Converts the tensor to a NumPy image using tensor2im
Constructs a save path based on the image name and visual label
Optionally resizes the image to match the desired aspect_ratio
Saves the image to disk using save_image
Collects all image paths, labels, and links for the gallery row

After processing all visuals, it calls webpage.add_images to insert the row into the HTML gallery.

Usage

Called in the test loop of test.py for each processed sample. The function is the primary interface between the model inference pipeline and the result visualization system.

Code Reference

Source Location

File	Lines
util/visualizer.py	L12-37
util/util.py	(tensor2im, save_image helper functions)

Signature

def save_images(webpage, visuals, image_path, aspect_ratio=1.0, width=256):
    """Save images to disk and add them to an HTML page.

    Parameters:
        webpage (html.HTML)             -- the HTML webpage object
        visuals (OrderedDict)           -- an ordered dict of images to display/save;
                                           keys are names, values are tensors
        image_path (str)                -- the file path of the current image sample
        aspect_ratio (float)            -- the aspect ratio of saved images (default: 1.0)
        width (int)                     -- the display width of images on the HTML page (default: 256)
    """

Helper functions (from util/util.py):

def tensor2im(input_image, imtype=np.uint8):
    """Convert a Tensor array into a numpy image array.

    Parameters:
        input_image (tensor) -- the input image tensor
        imtype (type)        -- the desired numpy dtype (default: np.uint8)
    Returns:
        image_numpy (numpy)  -- the converted numpy image array
    """

def save_image(image_numpy, image_path, aspect_ratio=1.0):
    """Save a numpy image to disk.

    Parameters:
        image_numpy (numpy) -- input numpy image array
        image_path (str)    -- the path of the image file
        aspect_ratio (float) -- aspect ratio for resizing
    """

Import

from util.visualizer import save_images
from util import html

I/O Contract

**Inputs**
Parameter	Type	Description
webpage	html.HTML	The HTML gallery object to add image rows to
visuals	OrderedDict	Dictionary mapping label names (str) to image tensors (torch.Tensor in [-1, 1])
image_path	str	File path of the current test sample (used to derive output filenames)
aspect_ratio	float	Aspect ratio for resizing saved images (default: 1.0, no resizing)
width	int	Display width in pixels for images on the HTML page (default: 256)

**Outputs**
Output	Type	Description
Image files	.png files	Saved to webpage.get_image_dir(), one per visual in the dict
HTML row	side effect	A new row of images added to the webpage object via add_images

Usage Examples

from util.visualizer import save_images
from util import html
from collections import OrderedDict

# Set up the webpage
web_dir = os.path.join(opt.results_dir, opt.name,
                       '{}_{}'.format(opt.phase, opt.epoch))
webpage = html.HTML(web_dir,
                    'Experiment = %s, Phase = %s, Epoch = %s' %
                    (opt.name, opt.phase, opt.epoch))

# In the test loop
for i, data in enumerate(dataset):
    model.set_input(data)
    model.test()
    visuals = model.get_current_visuals()  # OrderedDict of tensors
    img_path = model.get_image_paths()     # list of file paths

    save_images(webpage, visuals, img_path,
                aspect_ratio=opt.aspect_ratio, width=opt.display_winsize)

# After the loop, save the gallery
webpage.save()

# Direct use of helper functions
from util.util import tensor2im, save_image
import numpy as np

# Convert a model output tensor to a displayable image
fake_B_tensor = model.fake_B  # shape: (1, 3, 256, 256), range: [-1, 1]
fake_B_image = tensor2im(fake_B_tensor)  # shape: (256, 256, 3), dtype: uint8

# Save to disk
save_image(fake_B_image, './output/fake_B.png')

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment