write a Image object into std.in buffer about pyvips HOT 6 CLOSED

Hi again, do you mind if I close this issue? It's confusing having two open for almost the same thing. Let's continue discussion in the other one.

from pyvips.

Hello, unfortunately libvips does not support read and write from pipes. You need to write to a memory buffer, then write that to stdout.

The fastest way is as a simple C-style memory array. There's an example here of passing an image between vips, numpy and PIL:

https://github.com/jcupitt/pyvips/blob/master/examples/pil-numpy-pyvips.py

Does ffmpeg support piping uncompressed areas of memory like that? If not, I guess you'll want the image coded as a jpg. You can use write_to_buffer to make a string containing a jpg-coded image:

https://jcupitt.github.io/pyvips/vimage.html#pyvips.Image.write_to_buffer

from pyvips.

Thanks John, I was able to get it to work using PIL as the mechanism to write to std.in, I just had to read your jpeg buffer as bytes. See below:

image_vips = image_offset.smartcrop(self.output_raster_width, self.output_raster_height)
data = image_vips.write_to_buffer('.JPEG', Q=95)
image_bytes = PIL.Image.open(io.BytesIO(data))
image_bytes.save(p.stdin, 'JPEG')

I tested the script you provided. Fast indeed! So this begs the question (forgive me if it's a silly question), for anything Vips does at speed is it just using np under the hood? The reason I ask is, I benchmarked both Pillow and Vips doing the same operations of resize, offset, crop, and I actually found Vips to be much slower. The vips code I used is roughly as follows:

self.original_image = pyvips.Image.new_from_file(self.file_path)
self.original_image_width = self.original_image.width
self.original_image_height = self.original_image.height

for frame in range(100):
    image_resize = self.original_image.resize(zoom)
    image_offset = image_resize.copy(xoffset = int(x_total), yoffset = int(y_total))
    image_vips = image_offset.smartcrop(self.output_raster_width, self.output_raster_height)
    data = image_vips.write_to_buffer('.JPEG', Q=95)
    image_bytes = PIL.Image.open(io.BytesIO(data))
    image_bytes.save(p.stdin, 'JPEG')

Is it slow because it's reading from disk every time (It shouldn't be though, since the object is established at the beginning)? Are the cffi mappings not actually in-play when not utilizing the numpy arrays?

from pyvips.

smartcrop is very slow. Try just crop(left, top, width, height).

from pyvips.

libvips has its own pixel functions, it doesn't use numpy.

You have:

image_vips = image_offset.smartcrop(self.output_raster_width, self.output_raster_height)
data = image_vips.write_to_buffer('.JPEG', Q=95)
image_bytes = PIL.Image.open(io.BytesIO(data))
image_bytes.save(p.stdin, 'JPEG')

You're encoding jpg twice there, there's no need. Just do something like:

image_vips = image_offset.smartcrop(self.output_raster_width, self.output_raster_height)
data = image_vips.write_to_buffer('.JPEG', Q=95)
sys.stdout.write(data)

from pyvips.

Hey John, just played off this in the other thread :-)

from pyvips.