The idea is to show very low res images, e.g. 64px wide, BUT: without interpolation, until the real image is fully loaded.
I got this idea from the website of the Sanity CMS . But they don't use it anymore :/
Original image (here: ~583 KB):
PixelPreview image (here: ~1.3KB):
No interpolation/Nearest (here: ~1,6KB):
Bilinear (here: ~1,3KB):
Bicubic (here: ~1,3KB):
Lanczos (here: ~1,3KB):
The result with no interpolation looks interesting but might be a bit too much.
I am probably going to use lanczos.
Goal: Every time a new image is uploaded or an existing image is modified a
PixelPreview version of that image should be generated and stored in the same
directory with an appendix like -pixel-preview
.
Example: .../image.jpeg
-> .../image-pixel-preview.jpeg
I use a cloud function for that which can be easily "hooked onto" a google storage bucket belonging to the same project.
All cloud function endpoints are located within main.py
. You can test the
implementation locally using main_test.py
but have to include a valid
service-account.json
from GCP within the same directory.
If I reduce any image to a fixed width of let's say 64px
the aspect ratios
of the original image an the resulting 64px
-image might differ.
No problem with images like 1920x1080
:
- Original:
1920x1080
with aspect ratio16:9
- PixelPreview:
64x36
with aspect ratio16:9
Problematic with images like 3000x2000
:
- Original:
3000x2000
with aspect ratio3:2
- PixelPreview:
64x42
or64x43
with aspect ratios3.0476:2
or2.9767:2
respectively
Idea: Not only generate the PixelPreview but also replace the original image with a cropped version if needed.
Calculating the cropped size that suffices a given PIXEL_PREVIEW_WIDTH
:
import math
PIXEL_PREVIEW_WIDTH = 64
def get_snap_size(size):
# size given as (width, height) 2-tuple
crop_width = size[0]
crop_height = size[1]
# 1) Calculated floored height of the pixel image (may differ in aspect ratio)
pixel_height = math.floor(PIXEL_PREVIEW_WIDTH * (crop_height/crop_width))
# 2) Calculated full height with the pixel-image-ratio
crop_height = (pixel_height/PIXEL_PREVIEW_WIDTH) * crop_width
# 3) When that full height is a whole number -> finished
while int(crop_height) != crop_height:
crop_width -= 1
pixel_height = math.floor(PIXEL_PREVIEW_WIDTH * (crop_height / crop_width))
crop_height = (pixel_height / PIXEL_PREVIEW_WIDTH) * crop_width
# The resulting image will have an aspect ratio of 64:1 or 64:2 ... 64:100 or 64:101 ...
return (crop_width, int(crop_height))
I use the function get_resize_region
to calculate the PixelPreview size.
I use the function get_crop_region
to calculate the crop 4-tuple
(x0, y0, x1, y1)
which crops the image in a centered manner.
A way better variant of this Algorithm would be a version that sequentially tested the crop sizes ordered by how many pixels of the image would be lost! That way the algorithm will always produce the cropped size with the least possible lost image area.
First I need the possible crops sorted descendingly by cropped area: [(3000, 2000), (2999, 2000), (3000, 1999), (2998, 2000), (2999, 1999), (3000, 1998), (2997, 2000), (2998, 1999), ...]
.
To be honest I did not figure out yet how to implement this way of counting downwards in code in an efficient way.
So I just implemented a function generating all of these possible sizes up to a certain
cutoff
value (dx,dy < cutoff
) and sorting them: get_crop_size_options
in crop.py
.
This function also supports passing it the cutoff from the previous options-list called
prev_cutoff
so that no options will be checked duplicately.
The function suffices_pixel_ratio
just checks whether the PixelPreview of an image has
the same aspect ratio as the original image:
def suffices_pixel_ratio(size, pixel_width=64):
# size given as a (width, height) 2-tuple
scaling_factor = pixel_width/size[0]
pixel_height = scaling_factor * size[1]
return int(pixel_height) == pixel_height
The new version of the algorithm is certainly easier to understand:
def get_snap_size(size):
# size given as a (width, height) 2-tuple
if suffices_pixel_ratio(size):
return size
prev_cutoff = 0
current_cutoff = 10
while (prev_cutoff < max(size)):
size_options = get_crop_size_options(
size,
cutoff=current_cutoff,
prev_cutoff=prev_cutoff)
for size_option in size_options:
if suffices_pixel_ratio(size_option):
return size_option
current_cutoff += 10
prev_cutoff += 10
return (0, 0)
This new version takes significantly longer than the old one. In crop_playground.py
you can see how I compared their performance. For a large amount of images I would
probably store these computations.
Test samples (not super representativ): 100.000 random image sizes with width in
[400, 4000[
and height in [250, 2500[
.
Average performance of the old algorithm:
- Loss of image area:
~ 4.34%
- Time taken:
~ 0.0124ms
Average performance of the new algorithm:
- Loss of image area:
~ 2.56%
(optimal) - Time taken:
~ 0.701ms
So way less loss of image area but about 57 times slower ...
What to to about it? Right now, nothing! Why? When taking into account what the rest of the program does it is obvious that +0.7ms/image is not a big issue compared to the rest.
Opening a locally stored image with PIL
and storing a cropped version of it takes
> 120ms
on the same hardware. In addition to that I am fetching the source image
from a storage bucket as well as write into a storage bucket two times (in case the
image has to be cropped).
So optimization could definitely be relevant in another setup but not here.
The goal is to have a React component PixelImagePreview
that updates
really fast with the src
of the preview image and loads the actual
image after it has been mounted. The desired result: For the
loading time the pixel preview image is visible.
For the top white part of these slideshow images the chosen downsampling function lanczos is not ideal! I will have to look for a better one here.
There is probably a better way to asynchronously load
images with JS after the placeholder <img>
tag has been mounted in
the DOM. Message me if you want to tell me how to improve the
implementation ;)
I am using images for this slideshow instead of a pdf because every js/react pdf library I tried out had a significant bundle size as well as way more runtime overhead than using plain images. In addition to that I am using the same image slider on the whole page (blog-posts, gallery, slideshows) so if I were to use a pdf then I would have to build a more logic for the pdf slider.