For best results, you should (at least) set jpeg quality to 95 and use the "lanczos" filter. Obviously this will require significantly more storage and CPU time. From our experience, "cubic" might be 30% faster than "lanczos" on modern desktop and server CPUs. Keep in mind that you only need to create thumbnails once and then can enjoy them for the rest of your life.
--jpeg-quality value, -q value
set to 95 for high-quality thumbnails (25-100) (default: 90) [$PHOTOPRISM_JPEG_QUALITY]
--thumb-size value, -s value
default thumbnail size limit in pixels (720-3840) (default: 2048) [$PHOTOPRISM_THUMB_SIZE]
--thumb-limit value, -x value
on-demand thumbnail size limit in pixels (720-3840) (default: 3840) [$PHOTOPRISM_THUMB_LIMIT]
on-demand rendering of default thumbnails (high memory and cpu usage) [$PHOTOPRISM_THUMB_UNCACHED]
--thumb-filter value, -f value
resample filter (best to worst: blackman, lanczos, cubic, linear) (default: "lanczos") [$PHOTOPRISM_THUMB_FILTER]
You can limit the size of JPEG thumbnails using the
-s parameter when you run commands:
photoprism -s 720 start
The minimum size is 720px. This will render the following images:
500px is used for tiles in search results, the others are mostly needed for classification.
Size is still ~550kb with high quality (95). With lower JPEG quality (80), you'll get it down to ~100kb:
photoprism -s 720 -q 80 start
This page demonstrates and discusses the effects of JPEG compression: http://fotoforensics.com/tutorial-estq.php
Image classification obviously works better with sharp images, so it's possible you'll get less accurate labels with higher compression. Please share your experience.
If size limit is exceeded, for example because you use a large screen, originals will be used for displaying images in the frontend. This might result in the image displayed with wrong rotation if the browser doesn't rotate it automatically.
|fit_2560||2560||1600||lightbox / retina|
|fit_3840||3840||2400||lightbox / retina|
Bilinear interpolation takes a weighted average of the four neighborhood pixels to calculate its final interpolated value. The result is much smoother image than the original image. When all known pixel distances are equal, then the interpolated value is simply their sum divided by four. This technique performs interpolation in both directions, horizontal and vertical. This technique gives better result than nearest neighbor interpolation and take less computation time compared to bicubic interpolation.
Catmull-Rom is a local interpolating spline developed for computer graphics purposes. Its initial use was in design of curves and surfaces, and has recently been used in several applications. Catmull-Rom splines are a family of cubic interpolating splines formulated such that the tangent at each point is calculated using the previous and next point on the spline. The results are similar to ones produced by bicubic interpolation with regards to sharpness, but the Catmull-Rom reconstruction is clearly superior in smooth signal region.
The Lanczos interpolation function is a mathematical formula used to smoothly interpolate the value of a digital image between its samples. It maps each sample of the given image to a translated and scaled copy of the Lanczos kernel, which is a sinc function windowed by the central hump of a dilated sinc function. The sum of these translated and scaled kernels is then evaluated at the desired pixel. Lanczos interpolation has the best properties in terms of detail preservation and minimal generation of aliasing artifacts for geometric transformations not involving strong down sampling. However higher order Lanczos interpolation requires high computational time, which makes it unsuitable for most commercial software.
Blackman is a modification of Lanczos that has better control of ringing artifacts.
The same image resized from 600x400px to 150x100px using different resampling filters. From faster (lower quality) to slower (higher quality):