Image padding to fix an odd quirk

[adamgeddon1686]

So I discovered another odd quirk that I actually think might be quite easy to address. It seems that your software has a bit of an issue when the face is super closeup but this can be easily fixed using something like PIL to pad the image borders with black pixels, run the swap and then remove the black pixels afterwards. This is what I was thinking. I was going to run some tests after work today and I will report back what I find.
Here is how I implement it in my version of your facedancer function. You need to import PIL and define the variable padding to your liking. I am going to experiment with what the optimal setting for this padding and check to make sure there aren't any negative side effects.

def pad_in():
	target_files = os.listdir(temp_one)
	for target_filename in target_files:
		img_path = os.path.join(temp_one, target_filename)
		img_output = os.path.join(temp_two, os.path.basename(img_path))
		
		image = Image.open(img_path)
		width, height = image.size

		new_width = width + (2 * padding)
		new_height = height + (2 * padding)
		new_image = Image.new("RGB", (new_width, new_height), (0, 0, 0))

		x = padding + (new_width - width) // 2
		y = padding + (new_height - height) // 2
		new_image.paste(image, (x, y))

		new_image.save(img_output)

	shuffle_up()


def pad_out():
	target_files = os.listdir(temp_one)
	for target_filename in target_files:
		img_path = os.path.join(temp_one, target_filename)
		img_output = os.path.join(temp_two, os.path.basename(img_path))

		image = Image.open(img_path)
		padded_width, padded_height = image.size
		original_width = padded_width - (2 * padding)
		original_height = padded_height - (2 * padding)

		left = padding
		top = padding
		right = padded_width - padding
		bottom = padded_height - padding

		cropped_image = image.crop((left, top, right, bottom))
		cropped_image.save(img_output)


def face_dancer():
	global count_step, check_pause
	if count_step == 3:
		if input('Next Step: Swapping faces with FaceDancer...Continue (y/n): ') == 'y':
			RetinaFace = load_model(opt.retina_path, compile=False,
									custom_objects={'FPN': FPN,
													'SSH': SSH,
													'BboxHead': BboxHead,
													'LandmarkHead': LandmarkHead,
													'ClassHead': ClassHead})
			ArcFace = load_model(opt.arcface_path, compile=False)

			G = load_model(opt.facedancer_path, compile=False,
						   custom_objects={'AdaIN': AdaIN,
										   'AdaptiveAttention': AdaptiveAttention,
										   'InstanceNormalization': InstanceNormalization})

			target_files = os.listdir(temp_one)
			pad_in()
			for target_filename in target_files:
				img_path = os.path.join(temp_one, target_filename)
				img_output = os.path.join(temp_two, os.path.basename(img_path))
				run_inference(opt, swap_source, img_path,
							  RetinaFace, ArcFace, G, img_output)
			pad_out()
			if check_pause == 'true':
				pause()
			shuffle_up()

[adamgeddon1686]

For anyone else looking at this. Two things. First, I am just learning how to code. It is very likely there are better ways to do just about everything I post. Second, this is only a snippet. There are import statements file structures and other functions that I made up that aren't included because they are convoluted and confusing and mostly driven just towards my purposes. If anyone is interested in colabing, just message me and I can share the whole code. It is ridiculously long because, like I said, I am almost positive I am not doing everything correct. But it does work, so I am mostly happy with it.

[felixrosberg]

Hi, yes. Not unexpected with face detectors.

I will try remember to add an option for padding the images for the next update!