ausboss
a416420bfa
A see-through part of a picture - a cutout's missing background, an oval photo's corners, a hole - kept its stored colour blended over the fill. That colour is often black or the old background, so the model got a dark or light ring round the picture and Stitch Inpaint pasted part of it back. A pixel less than 90% as solid as the picture's most solid pixel is now empty, like a corner a turn leaves: filled with the fill colour, marked in the mask, new area for both seams. Kept pixels are fully solid in their own colour. A new last output, prompt_image, is the image with see-through parts on white instead of the fill, for the node that writes the prompt: shown the gray fill, a prompt writer describes a gray backdrop and the model keeps it. original shows see-through parts on white instead of black. A picture at least 90% solid everywhere goes through exactly as before; the new tests hold every output and both seams to the old transform.
260 lines
15 KiB
Python
260 lines
15 KiB
Python
from pathlib import Path
|
|
import sys
|
|
import os
|
|
import unittest
|
|
from unittest.mock import patch
|
|
|
|
import numpy as np
|
|
from PIL import Image
|
|
import torch
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
|
from nodes.node_image_crop_rotate_pad import AusBossImageCropRotatePad
|
|
from nodes.node_video_crop_rotate_pad import AusBossVideoCropRotatePad
|
|
from nodes._color_helpers import rgb_to_lab
|
|
from nodes._inpaint_crop_helpers import apply_stitch, stitch_blend_from_mask
|
|
|
|
COMFY_ROOT = os.environ.get('AUSBOSS_COMFY_ROOT')
|
|
if COMFY_ROOT:
|
|
sys.path.append(COMFY_ROOT)
|
|
|
|
|
|
class SeamColorMatchTests(unittest.TestCase):
|
|
"""color_match reads each padded seam off picture pixels only.
|
|
|
|
A rotated crop keeps its void corners inside the source rectangle, and
|
|
the seam reading used to count their fill as picture: a model that
|
|
continued a flat green picture perfectly came back up to 33 dE off.
|
|
"""
|
|
|
|
COLOUR = (46, 150, 72)
|
|
LAYOUTS = {
|
|
"rotated and padded": dict(rotation_degrees=-9, pad_left=24, pad_right=40),
|
|
"rotated to a canvas multiple": dict(rotation_degrees=17, canvas_multiple=64),
|
|
"rotated, cropped and padded": dict(
|
|
rotation_degrees=-18.1, crop_x=10, crop_y=6, crop_width=120, crop_height=100,
|
|
pad_left=40, pad_top=30, pad_right=40, pad_bottom=30,
|
|
),
|
|
}
|
|
|
|
def load(self, *frames, **values):
|
|
with patch('nodes.node_image_crop_rotate_pad.resolve_input_path', return_value=Path('source.png')), patch('nodes.node_image_crop_rotate_pad.load_image_frames', return_value=list(frames)):
|
|
return AusBossImageCropRotatePad().load_transform('source.png', feather=0, **values)
|
|
|
|
def fill_error(self, stitched, expected, mask):
|
|
from nodes._color_helpers import rgb_to_lab
|
|
|
|
error = (rgb_to_lab(stitched) - rgb_to_lab(expected)).norm(dim=-1)
|
|
fill = mask > 0.5
|
|
self.assertTrue(fill.any())
|
|
return float(error[fill].max()), float(error.max())
|
|
|
|
def test_a_perfect_continuation_keeps_its_fill(self):
|
|
source = Image.new('RGB', (128, 96), self.COLOUR)
|
|
colour = torch.tensor(self.COLOUR, dtype=torch.float32) / 255
|
|
for name, layout in self.LAYOUTS.items():
|
|
with self.subTest(layout=name):
|
|
image, mask, stitcher, *_ = self.load(source, **layout)
|
|
continuation = colour.expand_as(image).clone()
|
|
stitched = apply_stitch(stitcher, continuation, color_match=1.0)
|
|
fill_max, overall_max = self.fill_error(stitched, continuation, mask)
|
|
self.assertLess(fill_max, 0.5)
|
|
self.assertLess(overall_max, 1.0)
|
|
self.assertIs(stitcher['generated'], mask)
|
|
|
|
def test_a_drifted_fill_comes_back_to_the_picture(self):
|
|
source = Image.new('RGB', (128, 96), self.COLOUR)
|
|
colour = torch.tensor(self.COLOUR, dtype=torch.float32) / 255
|
|
for name, layout in self.LAYOUTS.items():
|
|
with self.subTest(layout=name):
|
|
image, mask, stitcher, *_ = self.load(source, **layout)
|
|
continuation = colour.expand_as(image).clone()
|
|
# The sampler continued the green 0.06 brighter everywhere it
|
|
# generated and mixed that into the band by the mask, as
|
|
# samplers do.
|
|
drifted = continuation + 0.06 * stitcher['blend'].unsqueeze(-1)
|
|
self.assertGreater(self.fill_error(apply_stitch(stitcher, drifted), continuation, mask)[0], 5.0)
|
|
stitched = apply_stitch(stitcher, drifted, color_match=1.0)
|
|
self.assertLess(self.fill_error(stitched, continuation, mask)[0], 1.0)
|
|
|
|
def test_the_source_stays_bit_identical_and_older_stitchers_still_stitch(self):
|
|
source = Image.fromarray(np.random.default_rng(5).integers(0, 256, (192, 256, 3), dtype=np.uint8))
|
|
image, mask, stitcher, *_ = self.load(source, **self.LAYOUTS["rotated and padded"])
|
|
generated = torch.rand(image.shape, generator=torch.Generator().manual_seed(6))
|
|
older = {key: value for key, value in stitcher.items() if key != 'generated'}
|
|
kept = stitcher['blend'] == 0
|
|
self.assertTrue(kept.any())
|
|
for current in (stitcher, older):
|
|
stitched = apply_stitch(current, generated, color_match=1.0)
|
|
self.assertTrue(torch.equal(stitched[kept], image[kept]))
|
|
|
|
def test_a_shorter_batch_trims_the_mask_with_the_blend(self):
|
|
# A video model hands back fewer frames than the stitcher holds.
|
|
source = Image.new('RGB', (128, 96), self.COLOUR)
|
|
image, mask, stitcher, *_ = self.load(source, source, source, **self.LAYOUTS["rotated and padded"])
|
|
self.assertEqual(stitcher['generated'].shape[0], 3)
|
|
colour = torch.tensor(self.COLOUR, dtype=torch.float32) / 255
|
|
continuation = colour.expand_as(image[:2]).clone()
|
|
stitched = apply_stitch(stitcher, continuation, color_match=1.0)
|
|
self.assertEqual(tuple(stitched.shape), tuple(continuation.shape))
|
|
self.assertLess(self.fill_error(stitched, continuation, mask[:2])[0], 0.5)
|
|
|
|
def test_load_image_pad_color_match_is_unchanged(self):
|
|
# Pinned: the published Qwen 2.1 Outpaint workflow stitches Load
|
|
# Image + Pad (feather 32) with color_match 1, and people who update
|
|
# the pack must get the same picture. These are the padded bands'
|
|
# mean colours from before the transform stitcher gained its mask.
|
|
from nodes.node_load_image_pad import AusBossLoadImagePad
|
|
|
|
y, x = np.mgrid[0:96, 0:128].astype(np.float32)
|
|
pattern = np.stack([
|
|
0.35 + 0.25 * np.sin(x / 9.0),
|
|
0.45 + 0.2 * np.cos(y / 7.0),
|
|
0.3 + 0.2 * np.sin((x + y) / 13.0),
|
|
], -1)
|
|
source = Image.fromarray((pattern * 255).round().astype(np.uint8))
|
|
with patch('nodes.node_load_image_pad.resolve_input_path', return_value=Path('source.png')), patch('nodes.node_load_image_pad.load_image_frames', return_value=[source]):
|
|
image, _, _, _, stitcher, _ = AusBossLoadImagePad().load_pad('source.png', 24, 16, 40, 8, 'color', '#808080', 0.5, 32, 8, 0.0)
|
|
self.assertNotIn('generated', stitcher)
|
|
rows = torch.arange(image.shape[1], dtype=torch.float32).view(1, -1, 1, 1)
|
|
drift = 0.06 + 0.04 * torch.sin(rows / 11.0)
|
|
inpainted = (image + drift * stitcher['blend'].unsqueeze(-1)).clamp(0, 1)
|
|
stitched = apply_stitch(stitcher, inpainted, color_match=1.0)
|
|
self.assertEqual(stitcher['source_bbox'], (24, 16, 152, 112))
|
|
bands = {
|
|
'left': (stitched[:, :, :24], (0.504521, 0.488853, 0.362618)),
|
|
'top': (stitched[:, :16], (0.453719, 0.476223, 0.33837)),
|
|
'right': (stitched[:, :, 152:], (0.471721, 0.495965, 0.318072)),
|
|
'bottom': (stitched[:, 112:], (0.435199, 0.536332, 0.34287)),
|
|
}
|
|
for side, (band, expected) in bands.items():
|
|
with self.subTest(side=side):
|
|
mean = band.double().mean(dim=(0, 1, 2))
|
|
self.assertTrue(torch.allclose(mean, torch.tensor(expected, dtype=torch.float64), atol=2e-6), mean)
|
|
|
|
|
|
class TransformOutputTests(unittest.TestCase):
|
|
def setUp(self):
|
|
self.source = Image.fromarray(np.arange(96 * 128 * 3, dtype=np.uint8).reshape(96, 128, 3))
|
|
|
|
def assert_outputs(self, outputs):
|
|
image, mask, stitcher, original, width, height = outputs[:6]
|
|
self.assertEqual((width, height), (160, 96))
|
|
expected = torch.from_numpy(np.asarray(self.source).astype(np.float32) / 255)
|
|
self.assertTrue(torch.equal(original[0], expected))
|
|
self.assertEqual(tuple(image.shape[1:3]), (96, 160))
|
|
self.assertEqual(tuple(mask.shape), (1, 96, 160))
|
|
generated = torch.ones_like(image)
|
|
stitched = apply_stitch(stitcher, generated)
|
|
kept = stitcher["blend"] == 0
|
|
self.assertTrue(kept.any())
|
|
self.assertTrue(torch.equal(stitched[kept], image[kept]))
|
|
self.assertTrue(torch.equal(stitched[:, :, -32:], image[:, :, -32:]))
|
|
|
|
def test_image_outputs_append_stitcher_and_untouched_source(self):
|
|
with patch('nodes.node_image_crop_rotate_pad.resolve_input_path', return_value=Path('source.png')), patch('nodes.node_image_crop_rotate_pad.load_image_frames', return_value=[self.source]):
|
|
outputs = AusBossImageCropRotatePad().load_transform('source.png', pad_left=32, feather=0)
|
|
self.assert_outputs(outputs)
|
|
# prompt_image, appended last: the image itself for an opaque picture.
|
|
self.assertEqual(len(outputs), 7)
|
|
self.assertTrue(torch.equal(outputs[6], outputs[0]))
|
|
self.assertIsNot(outputs[6], outputs[0])
|
|
|
|
def test_video_frame_has_the_same_outputs(self):
|
|
with patch('nodes.node_video_crop_rotate_pad.resolve_video_path', return_value=Path('source.mp4')), patch('nodes.node_video_crop_rotate_pad.decode_video_frame', return_value=(self.source, 0, 0)):
|
|
outputs = AusBossVideoCropRotatePad().load_transform('source.mp4', 'input folder', '', 'frame index', 0, 0, pad_left=32, feather=0)
|
|
self.assert_outputs(outputs)
|
|
|
|
def test_resize_inputs_are_optional_for_old_api_prompts(self):
|
|
inputs = AusBossImageCropRotatePad.INPUT_TYPES()
|
|
resize = ["resize_to_megapixels", "megapixels", "resize_method", "resolution_steps"]
|
|
# The resize inputs keep their widget positions; the stitch settings follow.
|
|
self.assertEqual(list(inputs["optional"]), resize + ["stitch_blend", "stitch_grow"])
|
|
self.assertFalse(set(resize) & set(inputs["required"]))
|
|
|
|
@unittest.skipUnless(COMFY_ROOT, "Set AUSBOSS_COMFY_ROOT for core resize integration")
|
|
def test_stitcher_uses_resized_canvas_dimensions(self):
|
|
with patch('nodes.node_image_crop_rotate_pad.resolve_input_path', return_value=Path('source.png')), patch('nodes.node_image_crop_rotate_pad.load_image_frames', return_value=[self.source]):
|
|
image, mask, stitcher, original, width, height, _ = AusBossImageCropRotatePad().load_transform('source.png', pad_left=32, feather=0, resize_to_megapixels=True, megapixels=.05, resolution_steps=16)
|
|
self.assertEqual((width, height), (int(image.shape[2]), int(image.shape[1])))
|
|
self.assertEqual(tuple(stitcher['canvas'].shape), tuple(image.shape))
|
|
self.assertEqual(tuple(original.shape), (1, 96, 128, 3))
|
|
self.assertEqual(tuple(apply_stitch(stitcher, image).shape), tuple(image.shape))
|
|
|
|
@unittest.skipUnless(COMFY_ROOT, "Set AUSBOSS_COMFY_ROOT for core resize integration")
|
|
def test_video_frame_resizes_to_a_budget_like_the_image_node(self):
|
|
with patch('nodes.node_video_crop_rotate_pad.resolve_video_path', return_value=Path('source.mp4')), patch('nodes.node_video_crop_rotate_pad.decode_video_frame', return_value=(self.source, 0, 0)):
|
|
image, mask, stitcher, original, width, height = AusBossVideoCropRotatePad().load_transform(
|
|
'source.mp4', 'input folder', '', 'frame index', 0, 0,
|
|
resize_to_megapixels=True, megapixels=.05, resolution_steps=16, pad_left=32, feather=0,
|
|
)
|
|
self.assertEqual((width, height), (int(image.shape[2]), int(image.shape[1])))
|
|
# 0.05 MP x 1024 x 1024 is larger than the 160 x 96 canvas: the
|
|
# budget scales the frame to it either way, aspect kept, to steps of 16.
|
|
budget = 0.05 * 1024 * 1024
|
|
self.assertLess(abs(width * height - budget) / budget, 0.1)
|
|
self.assertEqual((width % 16, height % 16), (0, 0))
|
|
self.assertEqual(tuple(stitcher['canvas'].shape), tuple(image.shape))
|
|
self.assertEqual(tuple(original.shape), (1, 96, 128, 3))
|
|
|
|
|
|
class FeatheredStitchTests(unittest.TestCase):
|
|
"""Feather shapes the mask; the canvas and the stitcher keep real pixels.
|
|
|
|
The transform used to fade the picture into the fill colour across the
|
|
feather ramp and hand that faded canvas to the stitcher, so color match
|
|
measured drift against it and pulled the fill toward grey.
|
|
"""
|
|
|
|
COLOUR = (46, 150, 72)
|
|
LAYOUTS = {
|
|
"straight pads": dict(pad_left=32, pad_top=24, pad_right=32, pad_bottom=24),
|
|
"rotated": dict(rotation_degrees=17),
|
|
"rotated and padded": dict(rotation_degrees=-9, pad_left=24, pad_right=40),
|
|
}
|
|
|
|
def load(self, source, **values):
|
|
with patch('nodes.node_image_crop_rotate_pad.resolve_input_path', return_value=Path('source.png')), patch('nodes.node_image_crop_rotate_pad.load_image_frames', return_value=[source]):
|
|
return AusBossImageCropRotatePad().load_transform('source.png', **values)
|
|
|
|
def test_color_match_keeps_the_fill_of_a_perfect_continuation(self):
|
|
# A model that continues a flat green picture perfectly paints the
|
|
# whole canvas that green; color match 1 must find no drift to fix.
|
|
source = Image.new('RGB', (128, 96), self.COLOUR)
|
|
colour = torch.tensor(self.COLOUR, dtype=torch.float32) / 255
|
|
# Rotated AND padded holds the rotation corners inside the crop; the
|
|
# seam reading skips them (SeamColorMatchTests), feathered or not.
|
|
for name, layout in self.LAYOUTS.items():
|
|
with self.subTest(layout=name):
|
|
image, mask, stitcher, *_ = self.load(source, feather=24, **layout)
|
|
continuation = colour.expand_as(image).clone()
|
|
stitched = apply_stitch(stitcher, continuation, color_match=1.0)
|
|
fill = mask > 0.5
|
|
error = (rgb_to_lab(stitched) - rgb_to_lab(continuation)).norm(dim=-1)
|
|
self.assertTrue(fill.any())
|
|
self.assertLess(float(error[fill].max()), 0.5)
|
|
self.assertLess(float(error.max()), 1.0)
|
|
|
|
def test_identity_round_trip_stays_exact(self):
|
|
source = Image.fromarray(np.random.default_rng(3).integers(0, 256, (96, 128, 3), dtype=np.uint8))
|
|
for name, layout in self.LAYOUTS.items():
|
|
with self.subTest(layout=name):
|
|
image, mask, stitcher, *_ = self.load(source, feather=24, **layout)
|
|
hard, *_ = self.load(source, feather=0, **layout)
|
|
# The feathered canvas is the unfeathered one, pixel for pixel.
|
|
self.assertTrue(torch.equal(image, hard))
|
|
self.assertTrue(torch.equal(apply_stitch(stitcher, image), image))
|
|
|
|
def test_stitch_blend_and_grow_reach_the_stitcher(self):
|
|
source = Image.new('RGB', (128, 96), self.COLOUR)
|
|
_, mask, stitcher, *_ = self.load(source, pad_left=32, feather=8)
|
|
self.assertTrue(torch.equal(stitcher['blend'], stitch_blend_from_mask(mask, 32, 0)))
|
|
_, mask, stitcher, *_ = self.load(source, pad_left=32, feather=8, stitch_blend=6, stitch_grow=-3)
|
|
self.assertTrue(torch.equal(stitcher['blend'], stitch_blend_from_mask(mask, 6, -3)))
|
|
_, mask, stitcher, *_ = self.load(source, pad_left=32, feather=0, stitch_blend=0)
|
|
self.assertTrue(torch.equal(stitcher['blend'], mask))
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|