overlay transparent image onto another image, but keep RGB channels

I am trying to overlay a watermark onto my image, such that the watermark is visible in the R,G,B channels but since the alpha channel is 0, the watermark is not visible on the overall image.

I created the watermark using convert watermark.png -alpha set -channel a -evaluate set 0% +channel watermark-t.png

I've spent a lot of time tweaking the following command, trying to use separate and -combine as well, but I am at my wit's end. Please help!

magick convert img.png watermark-t.png -gravity North-West -composite img_plus_watermark.png

When I use the above command, the watermark gets overlaid, but is not visible in any of the (including RGB) layers.

I've also tried using Pillow:

from PIL import Image
img = Image.open("img.png")
watermark = Image.open("watermark-t.png")

r, g, b, a = watermark.split()
watermark = Image.merge("RGB", (r, g, b))
mask = Image.merge("L", (a,))
img.paste(watermark, (0, 0), mask)
img.save("img_plus_watermark.png")

Still doesn't work.

The closest I came is to mask the img and watermark in the area where the watermark is overlaid. This leaves a transparent chunk of image in that area, but the image and watermark are visible in the RGB layers. I don't want this though.

Are PNGs just not equipped to handle this since they have only one alpha channel? Are there other image formats like TIFFs which contain layers that can do better -- if yes how?

Watermark:

Watermark original

Watermark transparent

Image

img+watermark-masked-to-remove-chunk-of-image

img+watermark-opaque

img+watermark-transparent (result of the above code)

If you're trying to hide a secret in the transparent parts of an image, that's actually pretty easy: remove the alpha from the image, paste in your solid RGB secret to the RGB parts of some transparent bit of the image, then reattach the original alpha.

For example, here's the standard PNG demo image -- the background is transparent:

And here's a tiny mario sprite (also an RGBA PNG with a transparent background):

I can remove the alpha from the demo image and the mario PNG like this (using libvips because I know it well, but I'm sure it'd be easy in imagemagick too):

$ vips extract_band PNG_transparency_demonstration.png bg.png 0 --n 3
$ vips extract_band mario.png fg.png 0 --n 3

That's extracting the bands (channels), starting at band 0 (red) and taking three of them (--n 3), so just the RGB parts.

Now I can paste mario in:

$ vips insert bg.png fg.png final-rgb.png 10 10

To make this RGB (no alpha) image:

Finally, I take the original alpha from the PNG demo and reattach it:

$ vips extract_band PNG_transparency_demonstration.png alpha.png 3
$ vips bandjoin "final-rgb.png alpha.png" result.png

To make this:

That looks like the PNG demo image, but it has a secret mario hidden in the top left corner. You need to remove the alpha to see it.

I am not sure I understand your issue. Just do the following in Imagemagick:

convert shoe.png \( watermark.png -alpha off \) -compose over -composite result.png

If this is not what you want, then please clarify further.

ADDITION

If you want the result to be transparent, but the watermark in the RGB channels, then you can use MPR: in memory storage to help to do that. It is like a named clone that can be used later by name.

convert shoe.png \( watermark.png -write mpr:img -alpha off \) \
-compose over -composite \( mpr:img -alpha extract \) \
-alpha off -compose copy_opacity -composite result2.png

convert result2.png -alpha off rgb.png

See https://imagemagick.org/Usage/files/#mpr

You cannot hide the watermark behind the image. Standard PNG images only contain one of each color band and one optional Alpha band (R, G, B, & A). So, there cannot be two RGBA values for any given pixel. In other words, when combining two PNGs the RGBA values of one image overwrite the other.

The alpha band (A) records the opacity of the RGB values in the image. So, you could set the alpha values of the watermark to 0 and add it to the image making it invisible (similar to what you've already done - see below).

from PIL import Image
img = Image.open("shoe.png")
watermark = Image.open("watermark-t.png")
watermark.putalpha(0)
img.paste(watermark)
img.save("img2.png")

Output:

You'll notice the top left corner is transparent and if the background was transparent too it would blend in but not in this case because it is white.

The code below will increase the alpha values of the image to show the watermark is still in the image.

from PIL import Image
img2 = Image.open("img2.png")
img2.putalpha(1000)
img2.save("img3.png")

Output:

Edit

Per your comment about the possibility of using tiff images:

I used the gdal library to generate a '7-band' tiff file that holds the color data for the watermark but does not display it (With the caveat that the watermark not being displayed may be dependent on the application used for displaying the image and the extra bands may cause some applications to be unable to read the image). I will add, this isn't the most efficient way to accomplish this but hopefully it is fairly clear.

If using Windows and installing gdal I would recommend downloading the wheel file of the library and using pip install <path to whl>.

from PIL import Image
from osgeo import gdal
# Convert shoe image from RGBA png to RGBA tif
img = Image.open("shoe.png")
img.save("shoe2.tif")
# Open the images
src_ds = gdal.OpenShared("shoe2.tif")
water_ds = gdal.OpenShared("watermark-t.png")
# Create memory copy of shoe tif
tmp_ds = gdal.GetDriverByName('MEM').CreateCopy('', src_ds, 0)
# loop 1,2,3 for gdal syntax when using GetRasterBand
for i in range(1,4):
    # Get watermark RGB bands as arrays
    wband = water_ds.GetRasterBand(i).ReadAsArray()
    # Add 5,6,7 bands to tif
    tmp_ds.AddBand()
    # Write RGB data to 5,6,7 bands of tif (4 is for Alpha)
    tmp_ds.GetRasterBand(i + 4).WriteArray(wband)
# Save as tif
dst_ds = gdal.GetDriverByName('GTiff').CreateCopy("img3.tif", tmp_ds, 0)
# Clean up
del src_ds
del water_ds
del tmp_ds
del dst_ds

The output (img3.tif) looks the same as shoe.png. Unfortunately it is too large to upload here (~14mb). To prove the watermark data is still intact in the tif the following code will take the 5, 6, & 7 bands and output them in a png.

from PIL import Image
from osgeo import gdal
# Create an empty png to store data in, size reflects the shoe image now
new_watermark = Image.new("RGB", (1938,1090))
new_watermark.save("empty.png")

img3_ds = gdal.OpenShared("img3.tif")
png_ds = gdal.OpenShared("empty.png")
water2_ds = gdal.GetDriverByName("MEM").CreateCopy("", png_ds, 0)
for i in range(1,4):
	band = img3_ds.GetRasterBand(i + 4).ReadAsArray()
	water2_ds.GetRasterBand(i).WriteArray(band)
output_ds = gdal.GetDriverByName("PNG").CreateCopy("remake_watermark.png", water2_ds, 0)
del img3_ds
del water2_ds
del output_ds

Output: