Adjusting image pixels internally in Blender with bpy

The naive approach

Tasked with representing some big data, let's see if blender can handle it. Here is some exploring first.

	import bpy
	D = bpy.data
	# BlendDataImages.new(name, width, height, alpha=False, float_buffer=False)
	image_object = D.images.new(name='pixeltest', width=4000, height=3000)
	# image_object points to the newly created: bpy.data.images['pixeltest']
	num_pixels = len(image_object.pixels)
	# well, turns out pixels is a little bit of a misnomer.
	print(num_pixels)
	# >>> 480000
	# A number 4 timeslarger than the expected 120000 pixels. It appears that
	# .pixels is a list of the RGBA values of every pixel in sequence.
	print( bpy.data.images['pixeltest'].file_format )
	# >>> 'TARGA'
	# seems to be the internal default
	print(image_object.pixels[:4])
	# >>> (0.0, 0.0, 0.0, 1.0)
	# drawing a pixel, changing pixel content
	for i in range(4):
	image_object.pixels[0+i] = (1.0)

view raw tmp.py hosted with ❤ by GitHub

adjusting the pixel on the last two lines above takes the most time. This image shows what the result is, zoomed in.

baby steps

This is relatively fast, but it's only 120 pixels in total. Try changing to 400*300 and you can expect it to take a lot longer, far too long to scale for big data.

	import bpy
	D = bpy.data
	image_object = D.images.new(name='pixeltest', width=40, height=30)
	num_pixels = len(image_object.pixels)
	# drawing a pixel, changing pixel content
	for px in range(0, num_pixels-5, 12):
	for i in range(4):
	image_object.pixels[px+i] = (1.0)

view raw pixmod_01.py hosted with ❤ by GitHub

	import bpy
	import random
	D = bpy.data
	image_object = D.images.new(name='pixeltest', width=40, height=30)
	num_pixels = len(image_object.pixels)
	# drawing a pixel, changing pixel content
	for px in range(0, num_pixels-5, 12):
	r = random.random()
	g = random.random()
	b = random.random()
	a = (1.0)
	cols = (r,g,b,a)
	for i in range(4):
	image_object.pixels[px+i] = cols[i]

view raw pixmod_02.py hosted with ❤ by GitHub

results in something profoundly uninteresting

If that isn't a good method, then perhaps construct the data and overwrite image_object in one go. You'll probably want to make sure the dimensions make sense.

What we know - end of naive

	>>> bpy.data.images['pixeltest'].pixels.__class__
	#<class 'bpy_prop_array'>

view raw pixel_array.py hosted with ❤ by GitHub

with a 40*30 image, i don't expect to notice much time difference but i'll know if the operation is possible.

dm = [(1.0) for i in range(4800)]
bpy.data.images['pixeltest'].pixels = dm
# turns them all white, so maybe try constructing the array first, then assigning.

This leads to a much faster way of pushing pixels. First create the image, then the array, then modify the array, then overwrite the image with the array data. The snippet below overwrites with a dark gray.

	import bpy
	import random
	D = bpy.data
	image_name = 'pixeltest2'
	if not image_name in D.images:
	image_object = D.images.new(name=image_name, width=400, height=300)
	image_object = D.images[image_name]
	num_pixels = len(image_object.pixels)
	dm = [(0.2) for cp in range(num_pixels)]
	image_object.pixels = dm

view raw faster_pixel_pushing.py hosted with ❤ by GitHub

And it seems that the speed is now closer to acceptable, here is a version that does a 400*300 px overwrite.

	import bpy
	import random
	D = bpy.data
	image_name = 'pixeltest3'
	if not image_name in D.images:
	image_object = D.images.new(name=image_name, width=400, height=300)
	image_object = D.images[image_name]
	num_pixels = len(image_object.pixels)
	dm = [(0.2) for cp in range(num_pixels)]
	for i in range(0,len(dm),16):
	for j in range(4):
	dm[i+j] = (1.0)
	image_object.pixels = dm

view raw lightning_fast.py hosted with ❤ by GitHub

4000*3000 will still take 10 seconds or so (on 2.4ghz 2core) but that's not too bad.

Great, what good is that to me?

I can think of a few applications but Firefly removal would be a top option. Find outliers and average the px values with the surrounding pixels

Finished firefly removal script can be found here: link