ausboss
ff59d55f41
An edit made at the picture's own size loses whatever the model pushed past the canvas edge, so realigning it left an empty strip on 13 of 16 held-out edits. Realign to Source now takes an optional stitcher: given Load Image + Pad's stitcher for the padded canvas the edit was made on, it measures inside the picture's area only and returns just that area, so what the model drew into the margin fills the edges. - A drift within half a pixel of a whole-pixel shift is cut straight out of the edit with no resampling; an edit that already lines up comes back untouched. - The coarse pass also tries the edit at 0.86x-1.28x, so a large zoom is found; models often reframe a padded canvas by zooming in to fill it, and that zoom does not count toward Max zoom. - The report says whether the frame was cut out or warped and names the margin each side needed when an edit ran out of picture. Load Image + Pad gains a mirror fill, which models read as more of the scene: over 16 edits it provoked a reframe on 2, against 7 for edge pixels and 5 for flat gray. The Qwen Image 2.1 Edit + Realign example now edits on a 64 px mirror margin with the stitcher wired.
811 lines
39 KiB
Python
811 lines
39 KiB
Python
from __future__ import annotations
|
|
|
|
from pathlib import Path
|
|
import sys
|
|
import unittest
|
|
|
|
import torch
|
|
|
|
ROOT = Path(__file__).resolve().parent.parent
|
|
sys.path.insert(0, str(ROOT))
|
|
if "nodes" in sys.modules and not hasattr(sys.modules["nodes"], "__path__"):
|
|
del sys.modules["nodes"]
|
|
|
|
import math
|
|
|
|
from nodes._pad_helpers import (
|
|
PAD_MODES,
|
|
_LOWRES_BLUR_MIN_SIGMA,
|
|
_SIGMA_DIVISOR,
|
|
_blur_image,
|
|
_resize_image,
|
|
feather_pad_mask,
|
|
pad_image,
|
|
plan_pad_canvas,
|
|
resolve_pad_geometry,
|
|
round_up_to_multiple,
|
|
)
|
|
|
|
|
|
def rand_image(batch: int, height: int, width: int, seed: int = 0) -> torch.Tensor:
|
|
generator = torch.Generator().manual_seed(seed)
|
|
return torch.rand((batch, height, width, 3), generator=generator, dtype=torch.float32)
|
|
|
|
|
|
PADS = (3, 5, 2, 4) # left, top, right, bottom
|
|
|
|
|
|
class MaskContractTests(unittest.TestCase):
|
|
def test_every_mode_shares_the_geometry_and_mask_contract(self):
|
|
image = rand_image(2, 16, 20, seed=1)
|
|
left, top, right, bottom = PADS
|
|
for mode in PAD_MODES:
|
|
with self.subTest(mode=mode):
|
|
out, mask = pad_image(image, left, top, right, bottom, mode)
|
|
self.assertEqual(out.shape, (2, 16 + top + bottom, 20 + left + right, 3))
|
|
self.assertEqual(mask.shape, (2, 16 + top + bottom, 20 + left + right))
|
|
# Original region: bit-identical image, zero mask.
|
|
self.assertTrue(
|
|
torch.equal(out[:, top : top + 16, left : left + 20, :], image)
|
|
)
|
|
inner = mask[:, top : top + 16, left : left + 20]
|
|
self.assertEqual(float(inner.sum()), 0.0)
|
|
# Padding: mask is exactly 1 everywhere else.
|
|
total = mask.numel() - inner.numel()
|
|
self.assertEqual(float(mask.sum()), float(total))
|
|
|
|
def test_zero_padding_is_a_passthrough_with_an_empty_mask(self):
|
|
image = rand_image(1, 8, 8, seed=2)
|
|
out, mask = pad_image(image, 0, 0, 0, 0, "edge")
|
|
self.assertTrue(torch.equal(out, image))
|
|
self.assertEqual(float(mask.sum()), 0.0)
|
|
|
|
def test_unknown_mode_is_rejected(self):
|
|
with self.assertRaises(ValueError):
|
|
pad_image(rand_image(1, 8, 8), 1, 1, 1, 1, "wrap")
|
|
|
|
|
|
class ColorModeTests(unittest.TestCase):
|
|
def test_padding_takes_the_parsed_fill_color(self):
|
|
image = rand_image(1, 8, 8, seed=3)
|
|
out, _ = pad_image(image, 2, 2, 2, 2, "color", fill_color="teal")
|
|
expected = torch.tensor([0.0, 128 / 255.0, 128 / 255.0])
|
|
self.assertTrue(torch.allclose(out[0, 0, 0], expected))
|
|
self.assertTrue(torch.allclose(out[0, -1, -1], expected))
|
|
self.assertTrue(torch.allclose(out[0, 0, 5], expected))
|
|
|
|
|
|
class EdgeModeTests(unittest.TestCase):
|
|
def test_sides_take_the_edge_average_and_corners_blend(self):
|
|
image = rand_image(1, 10, 12, seed=4)
|
|
left, top, right, bottom = PADS
|
|
out, _ = pad_image(image, left, top, right, bottom, "edge")
|
|
top_color = image[0, 0, :, :].mean(dim=0)
|
|
left_color = image[0, :, 0, :].mean(dim=0)
|
|
bottom_color = image[0, -1, :, :].mean(dim=0)
|
|
right_color = image[0, :, -1, :].mean(dim=0)
|
|
# Side bands are flat fills of the adjacent edge average.
|
|
self.assertTrue(torch.allclose(out[0, 0, left + 3], top_color))
|
|
self.assertTrue(torch.allclose(out[0, -1, left + 3], bottom_color))
|
|
self.assertTrue(torch.allclose(out[0, top + 3, 0], left_color))
|
|
self.assertTrue(torch.allclose(out[0, top + 3, -1], right_color))
|
|
# Corner quadrants blend their two adjoining sides.
|
|
self.assertTrue(torch.allclose(out[0, 0, 0], (top_color + left_color) / 2))
|
|
self.assertTrue(torch.allclose(out[0, -1, -1], (bottom_color + right_color) / 2))
|
|
|
|
|
|
class EdgePixelModeTests(unittest.TestCase):
|
|
def test_rows_and_cols_replicate_and_corners_take_the_corner_pixel(self):
|
|
image = rand_image(1, 10, 12, seed=5)
|
|
left, top, right, bottom = PADS
|
|
out, _ = pad_image(image, left, top, right, bottom, "edge pixel")
|
|
# Above the image, each column repeats the top source pixel of it.
|
|
for j in (0, 5, 11):
|
|
self.assertTrue(torch.equal(out[0, 0, left + j], image[0, 0, j]))
|
|
# Left of the image, each row repeats its leftmost source pixel.
|
|
for i in (0, 4, 9):
|
|
self.assertTrue(torch.equal(out[0, top + i, 0], image[0, i, 0]))
|
|
# Corner quadrants are the corner pixel.
|
|
self.assertTrue(torch.equal(out[0, 0, 0], image[0, 0, 0]))
|
|
self.assertTrue(torch.equal(out[0, 0, -1], image[0, 0, -1]))
|
|
self.assertTrue(torch.equal(out[0, -1, 0], image[0, -1, 0]))
|
|
self.assertTrue(torch.equal(out[0, -1, -1], image[0, -1, -1]))
|
|
|
|
|
|
class MirrorModeTests(unittest.TestCase):
|
|
def test_the_padding_mirrors_the_picture_at_every_edge(self):
|
|
image = rand_image(1, 10, 12, seed=6)
|
|
left, top, right, bottom = PADS
|
|
out, _ = pad_image(image, left, top, right, bottom, "mirror")
|
|
# Laid against the edge: the first padded row is the edge row itself,
|
|
# the next is the row after it, and so on outward.
|
|
for k in range(top):
|
|
self.assertTrue(torch.equal(out[0, top - 1 - k, left:left + 12], image[0, k]))
|
|
for k in range(bottom):
|
|
self.assertTrue(torch.equal(out[0, top + 10 + k, left:left + 12], image[0, 9 - k]))
|
|
for k in range(left):
|
|
self.assertTrue(torch.equal(out[0, top:top + 10, left - 1 - k], image[0, :, k]))
|
|
for k in range(right):
|
|
self.assertTrue(torch.equal(out[0, top:top + 10, left + 12 + k], image[0, :, 11 - k]))
|
|
# Corners mirror both ways.
|
|
self.assertTrue(torch.equal(out[0, 0, 0], image[0, top - 1, left - 1]))
|
|
|
|
def test_a_pad_wider_than_the_picture_keeps_reflecting(self):
|
|
image = rand_image(1, 3, 4, seed=7)
|
|
out, mask = pad_image(image, 9, 7, 10, 8, "mirror")
|
|
self.assertEqual(out.shape, (1, 3 + 7 + 8, 4 + 9 + 10, 3))
|
|
# Every padded pixel is some pixel of the picture.
|
|
colours = {tuple(c.tolist()) for c in image.reshape(-1, 3)}
|
|
for c in out.reshape(-1, 3):
|
|
self.assertIn(tuple(c.tolist()), colours)
|
|
# One period (flip + flip back) away along a row is the same pixel.
|
|
self.assertTrue(torch.equal(out[0, 7, 9 - 8], out[0, 7, 9]))
|
|
|
|
|
|
class PillarboxBlurModeTests(unittest.TestCase):
|
|
def test_backdrop_is_derived_from_the_image_not_a_flat_fill(self):
|
|
rows = torch.linspace(0.0, 1.0, 24).view(1, 24, 1, 1)
|
|
cols = torch.linspace(0.0, 1.0, 24).view(1, 1, 24, 1)
|
|
image = (0.5 * rows + 0.5 * cols).expand(1, 24, 24, 3).clone()
|
|
out, mask = pad_image(image, 12, 0, 12, 0, "pillarbox blur", backdrop_blur=0.5)
|
|
self.assertEqual(out.shape, (1, 24, 48, 3))
|
|
band = out[0, :, :12, :]
|
|
# The band varies (it is image content), and stays dimmer than the
|
|
# brightest source content because of the dim factor.
|
|
self.assertGreater(float(band.max() - band.min()), 0.05)
|
|
self.assertLessEqual(float(band.max()), 1.0 - 0.5 * 0.5 + 1e-4)
|
|
self.assertFalse(bool(out.isnan().any()))
|
|
self.assertEqual(float(mask[0, :, 12:36].sum()), 0.0)
|
|
|
|
def test_zero_strength_keeps_the_backdrop_sharp_and_undimmed(self):
|
|
image = rand_image(1, 16, 16, seed=6)
|
|
out, _ = pad_image(image, 8, 0, 8, 0, "pillarbox blur", backdrop_blur=0.0)
|
|
self.assertFalse(bool(out.isnan().any()))
|
|
# No dimming at strength 0: the padding can reach source brightness.
|
|
self.assertGreater(float(out[0, :, :8, :].max()), 0.5)
|
|
|
|
def test_stronger_setting_blurs_more(self):
|
|
generator = torch.Generator().manual_seed(7)
|
|
image = torch.rand((1, 32, 32, 3), generator=generator)
|
|
soft, _ = pad_image(image, 16, 0, 16, 0, "pillarbox blur", backdrop_blur=0.2)
|
|
hard, _ = pad_image(image, 16, 0, 16, 0, "pillarbox blur", backdrop_blur=1.0)
|
|
# More blur = less local variation in the backdrop band.
|
|
self.assertLess(
|
|
float(hard[0, :, :16, :].std()), float(soft[0, :, :16, :].std())
|
|
)
|
|
|
|
|
|
|
|
|
|
class LowResBackdropBlurTests(unittest.TestCase):
|
|
"""A heavy pillarbox blur runs at quarter resolution; a light one stays on
|
|
the exact full-resolution path."""
|
|
|
|
def test_heavy_blur_matches_the_full_resolution_reference(self):
|
|
torch.manual_seed(0)
|
|
image = torch.rand((2, 96, 160, 3))
|
|
out, _ = pad_image(image, 0, 120, 0, 120, "pillarbox blur", backdrop_blur=1.0)
|
|
# Rebuild the old full-resolution backdrop for the padded band.
|
|
canvas_h, canvas_w = 96 + 240, 160
|
|
scale = max(canvas_w / 160, canvas_h / 96)
|
|
sw = max(canvas_w, math.ceil(160 * scale))
|
|
sh = max(canvas_h, math.ceil(96 * scale))
|
|
backdrop = _resize_image(image, sw, sh)
|
|
cx, cy = (sw - canvas_w) // 2, (sh - canvas_h) // 2
|
|
backdrop = backdrop[:, cy : cy + canvas_h, cx : cx + canvas_w, :]
|
|
sigma = min(canvas_h, canvas_w) / _SIGMA_DIVISOR
|
|
self.assertGreaterEqual(sigma, _LOWRES_BLUR_MIN_SIGMA) # heavy path taken
|
|
reference = _blur_image(backdrop, sigma) * 0.5
|
|
band = out[:, :120, :, :]
|
|
self.assertLess(float((band - reference[:, :120, :, :]).abs().mean()), 0.01)
|
|
|
|
def test_light_blur_stays_on_the_exact_path(self):
|
|
torch.manual_seed(1)
|
|
image = torch.rand((1, 40, 64, 3))
|
|
# sigma = 0.2 * 56 / 16 = 0.7, far below the low-res threshold.
|
|
sigma = 0.2 * min(40 + 16, 64) / _SIGMA_DIVISOR
|
|
self.assertLess(sigma, _LOWRES_BLUR_MIN_SIGMA)
|
|
out, _ = pad_image(image, 0, 8, 0, 8, "pillarbox blur", backdrop_blur=0.2)
|
|
self.assertEqual(tuple(out.shape), (1, 56, 64, 3))
|
|
# The padded band still derives from the picture, not a flat fill.
|
|
band = out[:, :8, :, :]
|
|
self.assertGreater(float(band.std()), 0.0)
|
|
|
|
|
|
NO_TRIM = {"trim_left": 0, "trim_top": 0, "trim_right": 0, "trim_bottom": 0}
|
|
|
|
|
|
class ResolvePadGeometryTests(unittest.TestCase):
|
|
def test_both_sides_padded_keeps_the_remainder_on_the_far_side(self):
|
|
geometry = resolve_pad_geometry(10, 8, 1, 2, 3, 4, 16)
|
|
self.assertEqual(
|
|
geometry,
|
|
{"left": 1, "top": 2, "right": 5, "bottom": 6, "width": 16, "height": 16, **NO_TRIM},
|
|
)
|
|
|
|
def test_multiple_one_and_negatives_are_normalized(self):
|
|
geometry = resolve_pad_geometry(10, 8, -5, 0, 3, 0, 1)
|
|
self.assertEqual(
|
|
geometry,
|
|
{"left": 0, "top": 0, "right": 3, "bottom": 0, "width": 13, "height": 8, **NO_TRIM},
|
|
)
|
|
self.assertEqual(round_up_to_multiple(0, 8), 0)
|
|
self.assertEqual(round_up_to_multiple(1, 8), 8)
|
|
self.assertEqual(round_up_to_multiple(16, 8), 16)
|
|
|
|
def test_geometry_matches_the_frontend_pin(self):
|
|
# The same numbers are pinned in tests/pad_canvas.test.mjs — the JS
|
|
# mirror and this implementation must drift together or not at all.
|
|
geometry = resolve_pad_geometry(800, 600, 10, 20, 30, 40, 8)
|
|
self.assertEqual(
|
|
geometry,
|
|
{"left": 10, "top": 20, "right": 30, "bottom": 44, "width": 840, "height": 664, **NO_TRIM},
|
|
)
|
|
|
|
def test_one_padded_side_takes_the_whole_remainder(self):
|
|
# Pinned in tests/pad_canvas.test.mjs as well. 100 + 20 rounds to
|
|
# 128 at 16, 100 + 30 to 192 at 64; the other axis is a multiple.
|
|
cases = {
|
|
(100, 64, 20, 0, 0, 0, 16): {"left": 28, "top": 0, "right": 0, "bottom": 0, "width": 128, "height": 64},
|
|
(100, 64, 0, 0, 20, 0, 16): {"left": 0, "top": 0, "right": 28, "bottom": 0, "width": 128, "height": 64},
|
|
(128, 100, 0, 30, 0, 0, 64): {"left": 0, "top": 92, "right": 0, "bottom": 0, "width": 128, "height": 192},
|
|
(128, 100, 0, 0, 0, 30, 64): {"left": 0, "top": 0, "right": 0, "bottom": 92, "width": 128, "height": 192},
|
|
}
|
|
for args, expected in cases.items():
|
|
with self.subTest(args=args):
|
|
self.assertEqual(resolve_pad_geometry(*args), {**expected, **NO_TRIM})
|
|
|
|
def test_an_axis_nobody_padded_trims_the_source_instead_of_growing_a_strip(self):
|
|
# The published Krea 2 Outpaint case at source scale: top and bottom
|
|
# padded, the 1130 px width left alone. It used to grow to 1136 with a
|
|
# 6 px strip on the right; now the source loses 5 px on each side.
|
|
self.assertEqual(
|
|
resolve_pad_geometry(1130, 638, 0, 271, 0, 509, 16),
|
|
{
|
|
"left": 0, "top": 271, "right": 0, "bottom": 515, "width": 1120, "height": 1424,
|
|
"trim_left": 5, "trim_top": 0, "trim_right": 5, "trim_bottom": 0,
|
|
},
|
|
)
|
|
self.assertEqual(
|
|
resolve_pad_geometry(1000, 750, 0, 0, 0, 0, 64),
|
|
{
|
|
"left": 0, "top": 0, "right": 0, "bottom": 0, "width": 960, "height": 704,
|
|
"trim_left": 20, "trim_top": 23, "trim_right": 20, "trim_bottom": 23,
|
|
},
|
|
)
|
|
# An odd trim puts the extra pixel on the far side, like Align Image.
|
|
geometry = resolve_pad_geometry(1001, 64, 0, 0, 0, 0, 16)
|
|
self.assertEqual((geometry["trim_left"], geometry["trim_right"], geometry["width"]), (4, 5, 992))
|
|
|
|
def test_a_source_smaller_than_one_multiple_still_grows(self):
|
|
# Nothing to trim down to: the width grows on the far side, while the
|
|
# unpadded height above one multiple is trimmed as usual.
|
|
self.assertEqual(
|
|
resolve_pad_geometry(10, 750, 0, 0, 0, 0, 16),
|
|
{
|
|
"left": 0, "top": 0, "right": 6, "bottom": 0, "width": 16, "height": 736,
|
|
"trim_left": 0, "trim_top": 7, "trim_right": 0, "trim_bottom": 7,
|
|
},
|
|
)
|
|
|
|
def test_the_remainder_never_lands_on_an_edge_nobody_padded(self):
|
|
import random
|
|
|
|
rng = random.Random(7)
|
|
for _ in range(3000):
|
|
size = rng.randint(1, 3000)
|
|
multiple = rng.choice([1, 8, 16, 64, rng.randint(2, 128)])
|
|
before = rng.choice([0, 0, rng.randint(1, 500)])
|
|
after = rng.choice([0, 0, rng.randint(1, 500)])
|
|
geometry = resolve_pad_geometry(size, 64, before, 0, after, 0, multiple)
|
|
left, right = geometry["left"], geometry["right"]
|
|
trim_left, trim_right = geometry["trim_left"], geometry["trim_right"]
|
|
with self.subTest(size=size, multiple=multiple, before=before, after=after):
|
|
self.assertEqual(geometry["width"] % multiple, 0)
|
|
self.assertEqual(size - trim_left - trim_right + left + right, geometry["width"])
|
|
self.assertGreaterEqual(left, before)
|
|
self.assertGreaterEqual(right, after)
|
|
if before == 0 and left:
|
|
self.fail("the remainder landed on an unpadded left edge")
|
|
if after == 0 and right and (before or size >= multiple):
|
|
self.fail("the remainder landed on an unpadded right edge")
|
|
if before or after:
|
|
self.assertEqual((trim_left, trim_right), (0, 0))
|
|
else:
|
|
self.assertLess(trim_left + trim_right, multiple)
|
|
self.assertIn(trim_right - trim_left, (0, 1))
|
|
|
|
|
|
class PlanPadCanvasTests(unittest.TestCase):
|
|
def test_target_off_is_a_passthrough(self):
|
|
plan = plan_pad_canvas(800, 600, 10, 20, 30, 40, 8, 0.0)
|
|
self.assertEqual(plan["scale"], 1.0)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (800, 600))
|
|
self.assertEqual((plan["width"], plan["height"]), (840, 664))
|
|
|
|
def test_megapixel_target_rescales_the_source_first(self):
|
|
# Pinned against tests/pad_canvas.test.mjs finalOutputSize.
|
|
plan = plan_pad_canvas(800, 600, 10, 20, 30, 40, 8, 1.0)
|
|
self.assertAlmostEqual(plan["scale"], 1.3389868666385072, places=12)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (1071, 803))
|
|
self.assertEqual((plan["width"], plan["height"]), (1128, 888))
|
|
# The plan lands within multiple-rounding distance of the target.
|
|
self.assertLess(abs(plan["width"] * plan["height"] / 1e6 - 1.0), 0.02)
|
|
# Everything still rounds to the multiple after scaling.
|
|
self.assertEqual(plan["width"] % 8, 0)
|
|
self.assertEqual(plan["height"] % 8, 0)
|
|
# The source was scaled, and the scaled pads add up to the canvas.
|
|
self.assertEqual(plan["source_width"] + plan["left"] + plan["right"], plan["width"])
|
|
self.assertEqual(plan["source_height"] + plan["top"] + plan["bottom"], plan["height"])
|
|
|
|
def test_target_off_trims_an_axis_nobody_padded(self):
|
|
plan = plan_pad_canvas(1130, 638, 0, 271, 0, 509, 16, 0.0)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (1130, 638))
|
|
self.assertEqual((plan["trim_left"], plan["trim_right"]), (5, 5))
|
|
self.assertEqual((plan["left"], plan["right"], plan["width"]), (0, 0, 1120))
|
|
|
|
def test_published_krea2_outpaint_canvas_has_no_strip(self):
|
|
# The Krea 2 Outpaint workflow as published: a 2720x1536 photo padded
|
|
# top and bottom at 16 and 1.6 MP. It used to place the source 1130
|
|
# wide on an 1136 canvas, a 6 px strip on the right that nobody
|
|
# padded. The width is now resized onto the multiple and sets the
|
|
# scale for the height, so the photo spans the canvas at its own
|
|
# shape. Pinned in tests/pad_canvas.test.mjs too.
|
|
plan = plan_pad_canvas(2720, 1536, 0, 652, 0, 1212, 16, 1.6)
|
|
self.assertEqual(plan["scale"], 1136 / 2720)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (1136, 642))
|
|
self.assertEqual((plan["left"], plan["right"], plan["width"]), (0, 0, 1136))
|
|
self.assertEqual((plan["top"], plan["bottom"], plan["height"]), (272, 510, 1424))
|
|
self.assertEqual(
|
|
(plan["trim_left"], plan["trim_top"], plan["trim_right"], plan["trim_bottom"]),
|
|
(0, 0, 0, 0),
|
|
)
|
|
# The same at 64: still one scale for both axes.
|
|
plan = plan_pad_canvas(2720, 1536, 0, 652, 0, 1212, 64, 1.6)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (1088, 614))
|
|
self.assertEqual((plan["left"], plan["right"], plan["width"], plan["height"]), (0, 0, 1088, 1408))
|
|
|
|
def test_one_padded_side_takes_the_remainder_after_the_resize(self):
|
|
# A photo padded top and right: the height's remainder joins the top
|
|
# now, where it used to add a 4 px strip along the unpadded bottom.
|
|
plan = plan_pad_canvas(502, 634, 0, 240, 320, 0, 16, 1.6)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (742, 937))
|
|
self.assertEqual((plan["top"], plan["bottom"], plan["height"]), (359, 0, 1296))
|
|
self.assertEqual((plan["left"], plan["right"], plan["width"]), (0, 474, 1216))
|
|
plan = plan_pad_canvas(502, 634, 0, 240, 320, 0, 64, 1.6)
|
|
self.assertEqual((plan["top"], plan["bottom"], plan["height"]), (415, 0, 1344))
|
|
self.assertEqual((plan["left"], plan["right"], plan["width"]), (0, 481, 1216))
|
|
|
|
def test_a_side_under_half_a_multiple_is_not_snapped(self):
|
|
# 20 px tall at 64: snapping it up to one multiple would more than
|
|
# double the scale and the padded width with it. It keeps the budget
|
|
# scale and grows on the far side instead.
|
|
plan = plan_pad_canvas(2000, 20, 100, 0, 100, 0, 64, 0.25)
|
|
self.assertEqual(plan["scale"], math.sqrt(0.25e6 / (2240 * 64)))
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (2641, 26))
|
|
self.assertEqual((plan["top"], plan["bottom"], plan["height"]), (0, 38, 64))
|
|
self.assertEqual(plan["width"], 2944)
|
|
|
|
def test_no_padding_snaps_each_side_like_align_image_resize(self):
|
|
plan = plan_pad_canvas(1000, 750, 0, 0, 0, 0, 16, 1.0)
|
|
self.assertEqual((plan["source_width"], plan["source_height"]), (1152, 864))
|
|
self.assertEqual((plan["width"], plan["height"]), (1152, 864))
|
|
self.assertEqual((plan["left"], plan["top"], plan["right"], plan["bottom"]), (0, 0, 0, 0))
|
|
|
|
def test_target_on_never_leaves_a_strip_or_trim_on_an_unpadded_axis(self):
|
|
import random
|
|
|
|
rng = random.Random(11)
|
|
for _ in range(2000):
|
|
width, height = rng.randint(16, 4000), rng.randint(16, 4000)
|
|
multiple = rng.choice([8, 16, 64])
|
|
pads = [rng.choice([0, 0, rng.randint(40, 1500)]) for _ in range(4)]
|
|
target = rng.choice([0.5, 1.0, 1.6, 4.0])
|
|
plan = plan_pad_canvas(width, height, *pads, multiple, target)
|
|
left, top, right, bottom = pads
|
|
with self.subTest(size=(width, height), pads=pads, multiple=multiple, target=target):
|
|
self.assertEqual(plan["width"] % multiple, 0)
|
|
self.assertEqual(plan["height"] % multiple, 0)
|
|
snapped = []
|
|
for size, before, after, pad_before, pad_after in (
|
|
(plan["source_width"], plan["left"], plan["right"], left, right),
|
|
(plan["source_height"], plan["top"], plan["bottom"], top, bottom),
|
|
):
|
|
if pad_before + pad_after:
|
|
continue
|
|
if size % multiple == 0:
|
|
snapped.append(True)
|
|
self.assertEqual((before, after), (0, 0))
|
|
else:
|
|
# Under half a multiple at the budget scale: left
|
|
# alone and grown on the far side.
|
|
self.assertLess(size, multiple)
|
|
self.assertEqual(before, 0)
|
|
self.assertEqual(
|
|
(plan["trim_left"], plan["trim_top"], plan["trim_right"], plan["trim_bottom"]),
|
|
(0, 0, 0, 0),
|
|
)
|
|
if snapped and (left + right == 0) != (top + bottom == 0):
|
|
# One unpadded axis sets a single scale: no stretch
|
|
# beyond the other side's half-pixel rounding.
|
|
aspect = width / height
|
|
self.assertAlmostEqual(
|
|
plan["source_width"] / plan["source_height"],
|
|
aspect,
|
|
delta=aspect / min(plan["source_width"], plan["source_height"]),
|
|
)
|
|
|
|
|
|
class FeatherPadMaskTests(unittest.TestCase):
|
|
def build_mask(self):
|
|
_, mask = pad_image(rand_image(1, 16, 20, seed=9), 3, 5, 2, 4, "color")
|
|
return mask # canvas 25 x 25, image at (3, 5) sized 20 x 16
|
|
|
|
def test_zero_feather_returns_the_mask_untouched(self):
|
|
mask = self.build_mask()
|
|
self.assertTrue(torch.equal(feather_pad_mask(mask, 3, 5, 2, 4, 0), mask))
|
|
|
|
def test_ramp_runs_inward_and_padding_stays_solid(self):
|
|
mask = self.build_mask()
|
|
out = feather_pad_mask(mask, 3, 5, 2, 4, 3)
|
|
# Padding is untouched: fully 1 in the bands and corners.
|
|
self.assertEqual(float(out[0, 0, 0]), 1.0)
|
|
self.assertEqual(float(out[0, 2, 12]), 1.0)
|
|
self.assertEqual(float(out[0, -1, -1]), 1.0)
|
|
# Top ramp descends 0.75 / 0.5 / 0.25 into the image (col 12 is clear
|
|
# of the left/right ramps).
|
|
self.assertAlmostEqual(float(out[0, 5, 12]), 0.75, places=5)
|
|
self.assertAlmostEqual(float(out[0, 6, 12]), 0.5, places=5)
|
|
self.assertAlmostEqual(float(out[0, 7, 12]), 0.25, places=5)
|
|
self.assertEqual(float(out[0, 8, 12]), 0.0)
|
|
# Left ramp likewise, on a row clear of the top/bottom ramps.
|
|
self.assertAlmostEqual(float(out[0, 10, 3]), 0.75, places=5)
|
|
self.assertAlmostEqual(float(out[0, 10, 5]), 0.25, places=5)
|
|
self.assertEqual(float(out[0, 10, 6]), 0.0)
|
|
# Bottom ramp ascends back toward the padding.
|
|
self.assertAlmostEqual(float(out[0, 20, 12]), 0.75, places=5)
|
|
self.assertAlmostEqual(float(out[0, 18, 12]), 0.25, places=5)
|
|
# The far interior stays clean.
|
|
self.assertEqual(float(out[0, 12, 12]), 0.0)
|
|
|
|
def test_unpadded_sides_are_not_feathered(self):
|
|
_, mask = pad_image(rand_image(1, 16, 20, seed=10), 0, 0, 0, 6, "color")
|
|
out = feather_pad_mask(mask, 0, 0, 0, 6, 4)
|
|
# Only the bottom ramp exists: top rows and side columns stay 0.
|
|
self.assertEqual(float(out[0, 0, :].sum()), 0.0)
|
|
self.assertEqual(float(out[0, 2:12, 0].sum()), 0.0)
|
|
self.assertGreater(float(out[0, 15, 10]), 0.0)
|
|
|
|
def test_ramp_width_is_capped_by_the_image_dimension(self):
|
|
mask = self.build_mask()
|
|
out = feather_pad_mask(mask, 3, 5, 2, 4, 500)
|
|
self.assertEqual(out.shape, mask.shape)
|
|
# The whole image column is ramped but values stay in (0, 1].
|
|
self.assertGreater(float(out[0, 12, 12]), 0.0)
|
|
self.assertLessEqual(float(out.max()), 1.0)
|
|
|
|
|
|
class FeatherCornerTests(unittest.TestCase):
|
|
"""Two padded sides meet without a crease: the ramps unite instead of
|
|
taking their maximum, so the corner is bilinear."""
|
|
|
|
def test_corner_is_the_union_of_the_two_ramps(self):
|
|
_, mask = pad_image(rand_image(1, 16, 20, seed=11), 3, 5, 2, 4, "color")
|
|
out = feather_pad_mask(mask, 3, 5, 2, 4, 3)
|
|
# Row 5 is the first image row (top ramp 0.75), col 3 the first image
|
|
# column (left ramp 0.75): union 1 - 0.25 * 0.25, not max 0.75.
|
|
self.assertAlmostEqual(float(out[0, 5, 3]), 1.0 - 0.25 * 0.25, places=5)
|
|
self.assertAlmostEqual(float(out[0, 6, 4]), 1.0 - 0.5 * 0.5, places=5)
|
|
# Along a side the other ramp is 0, so nothing changes there.
|
|
self.assertAlmostEqual(float(out[0, 5, 12]), 0.75, places=5)
|
|
self.assertAlmostEqual(float(out[0, 10, 3]), 0.75, places=5)
|
|
|
|
def test_corner_has_no_diagonal_crease(self):
|
|
# With max(), stepping across the corner's diagonal flips which ramp
|
|
# wins and the gradient jumps. With the union the value along the
|
|
# diagonal is symmetric and the differences to its two neighbours
|
|
# are equal - no crease to print.
|
|
_, mask = pad_image(rand_image(1, 60, 60, seed=12), 20, 20, 20, 20, "color")
|
|
out = feather_pad_mask(mask, 20, 20, 20, 20, 16)
|
|
for step in range(1, 15):
|
|
y, x = 20 + step, 20 + step
|
|
self.assertAlmostEqual(float(out[0, y, x - 1]), float(out[0, y - 1, x]), places=5)
|
|
self.assertLess(float(out[0, y, x]), float(out[0, y, x - 1]))
|
|
|
|
|
|
class LoadImagePadNodeTests(unittest.TestCase):
|
|
def make_node(self):
|
|
from nodes.node_load_image_pad import (
|
|
NODE_CLASS_MAPPINGS,
|
|
NODE_DISPLAY_NAME_MAPPINGS,
|
|
)
|
|
|
|
self.assertIn("AUSBOSS_NODES_LoadImagePad", NODE_CLASS_MAPPINGS)
|
|
self.assertEqual(
|
|
NODE_DISPLAY_NAME_MAPPINGS["AUSBOSS_NODES_LoadImagePad"],
|
|
"Load Image + Pad 🆎",
|
|
)
|
|
return NODE_CLASS_MAPPINGS["AUSBOSS_NODES_LoadImagePad"]
|
|
|
|
def write_image(self, directory: str, width: int = 64, height: int = 48) -> str:
|
|
from PIL import Image
|
|
|
|
path = Path(directory) / "source.png"
|
|
Image.new("RGB", (width, height), (30, 180, 90)).save(path)
|
|
return str(path)
|
|
|
|
def run_node(self, cls, image_path: str, **overrides):
|
|
values = {
|
|
"pad_left": 2,
|
|
"pad_top": 3,
|
|
"pad_right": 4,
|
|
"pad_bottom": 5,
|
|
"mode": "color",
|
|
"fill_color": "#000000",
|
|
"backdrop_blur": 0.5,
|
|
"feather": 0,
|
|
"canvas_multiple": 8,
|
|
"target_megapixels": 0.0,
|
|
}
|
|
values.update(overrides)
|
|
return getattr(cls(), cls.FUNCTION)(image=image_path, **values), values
|
|
|
|
def test_contract_and_plain_padding(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
self.assertIn("AusBoss/Image", cls.CATEGORY)
|
|
self.assertEqual(
|
|
cls.RETURN_TYPES,
|
|
("IMAGE", "MASK", "INT", "INT", "AUSBOSS_STITCHER", "IMAGE"),
|
|
)
|
|
self.assertEqual(
|
|
cls.RETURN_NAMES,
|
|
("image", "mask", "width", "height", "stitcher", "reference"),
|
|
)
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(image, mask, width, height, _, _), _ = self.run_node(cls, path)
|
|
self.assertEqual((width, height), (72, 56)) # 70x56 ceiled to 8
|
|
self.assertEqual(tuple(image.shape), (1, 56, 72, 3))
|
|
self.assertEqual(tuple(mask.shape), (1, 56, 72))
|
|
# Hard mask, and the original lands intact at (2, 3).
|
|
self.assertEqual(set(torch.unique(mask).tolist()), {0.0, 1.0})
|
|
self.assertEqual(float(mask[0, 3 : 3 + 48, 2 : 2 + 64].sum()), 0.0)
|
|
|
|
def test_megapixel_target_resizes_source_first_and_keeps_the_seam_hard(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(image, mask, width, height, _, _), values = self.run_node(
|
|
cls, path, target_megapixels=0.05
|
|
)
|
|
plan = plan_pad_canvas(64, 48, 2, 3, 4, 5, 8, 0.05)
|
|
self.assertEqual((width, height), (plan["width"], plan["height"]))
|
|
self.assertEqual(tuple(image.shape), (1, height, width, 3))
|
|
self.assertLess(abs(width * height / 1e6 - 0.05), 0.01)
|
|
# Source-first order: the mask's zero region is exactly the
|
|
# RESIZED source rect, and the seam is still binary because the
|
|
# padding happened after the resize.
|
|
self.assertEqual(set(torch.unique(mask).tolist()), {0.0, 1.0})
|
|
inner = mask[
|
|
0,
|
|
plan["top"] : plan["top"] + plan["source_height"],
|
|
plan["left"] : plan["left"] + plan["source_width"],
|
|
]
|
|
self.assertEqual(float(inner.sum()), 0.0)
|
|
self.assertEqual(
|
|
float(mask.sum()), float(mask.numel() - inner.numel())
|
|
)
|
|
|
|
def test_feather_softens_the_seam_but_not_the_padding(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(_, mask, _, _, _, _), _ = self.run_node(cls, path, feather=6)
|
|
values = torch.unique(mask).tolist()
|
|
self.assertTrue(any(0.0 < value < 1.0 for value in values))
|
|
self.assertEqual(float(mask[0, 0, 0]), 1.0) # padding stays solid
|
|
|
|
def test_reference_is_the_unpadded_source_snapped_to_16(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(_, _, _, _, _, reference), _ = self.run_node(cls, path)
|
|
# 64x48 is already under the 384 cap and already /16.
|
|
self.assertEqual(tuple(reference.shape), (1, 48, 64, 3))
|
|
# It is the SOURCE, not the canvas: no padding colour in it.
|
|
self.assertGreater(float(reference.min()), 0.0)
|
|
|
|
def test_reference_downscales_a_large_source_to_a_multiple_of_16(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp, width=1000, height=750)
|
|
(_, _, _, _, _, reference), _ = self.run_node(cls, path)
|
|
_batch, height, width, _channels = reference.shape
|
|
self.assertLessEqual(max(width, height), 384)
|
|
self.assertEqual(width % 16, 0)
|
|
self.assertEqual(height % 16, 0)
|
|
|
|
def test_stitcher_carries_where_the_source_sits(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(_, mask, width, height, stitcher, _), _ = self.run_node(cls, path)
|
|
# pad_left=2, pad_top=3 on a 64x48 source.
|
|
self.assertEqual(stitcher["source_bbox"], (2, 3, 66, 51))
|
|
x0, y0, x1, y1 = stitcher["bbox_normalized"]
|
|
self.assertAlmostEqual(x0, 2 / width)
|
|
self.assertAlmostEqual(y0, 3 / height)
|
|
self.assertAlmostEqual(x1, 66 / width)
|
|
self.assertAlmostEqual(y1, 51 / height)
|
|
# The bbox is exactly the region the mask protects.
|
|
inner = mask[0, 3:51, 2:66]
|
|
self.assertEqual(float(inner.sum()), 0.0)
|
|
self.assertEqual(float(mask.sum()), float(mask.numel() - inner.numel()))
|
|
|
|
def test_bbox_tracks_the_megapixel_resize(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(_, mask, _, _, stitcher, _), _ = self.run_node(
|
|
cls, path, target_megapixels=0.05
|
|
)
|
|
plan = plan_pad_canvas(64, 48, 2, 3, 4, 5, 8, 0.05)
|
|
x0, y0, x1, y1 = stitcher["source_bbox"]
|
|
self.assertEqual((x0, y0), (plan["left"], plan["top"]))
|
|
self.assertEqual(x1 - x0, plan["source_width"])
|
|
self.assertEqual(y1 - y0, plan["source_height"])
|
|
self.assertEqual(float(mask[0, y0:y1, x0:x1].sum()), 0.0)
|
|
|
|
def test_an_edge_nobody_padded_stays_photo_through_the_stitch(self):
|
|
import tempfile
|
|
|
|
from PIL import Image
|
|
|
|
from nodes._inpaint_crop_helpers import apply_stitch
|
|
|
|
cls = self.make_node()
|
|
generator = torch.Generator().manual_seed(21)
|
|
pixels = (torch.rand((50, 70, 3), generator=generator) * 255).to(torch.uint8)
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = str(Path(tmp) / "source.png")
|
|
Image.fromarray(pixels.numpy()).save(path)
|
|
source = pixels.float().unsqueeze(0) / 255.0
|
|
for target in (0.0, 0.01):
|
|
with self.subTest(target_megapixels=target):
|
|
(image, mask, width, height, stitcher, _), _ = self.run_node(
|
|
cls, path, pad_left=0, pad_top=6, pad_right=0, pad_bottom=10,
|
|
canvas_multiple=16, target_megapixels=target,
|
|
)
|
|
self.assertEqual(width % 16, 0)
|
|
# The source spans the whole width: no strip for the model
|
|
# to paint, no seam for tone match to read on either side.
|
|
x0, y0, x1, y1 = stitcher["source_bbox"]
|
|
self.assertEqual((x0, x1), (0, width))
|
|
self.assertEqual(float(mask[0, y0:y1].sum()), 0.0)
|
|
if target == 0.0:
|
|
# 70 px trims to 64, three columns off each side, and
|
|
# what is kept lands bit-identical.
|
|
self.assertEqual(width, 64)
|
|
self.assertTrue(torch.equal(image[:, y0:y1], source[:, :, 3:67]))
|
|
# Whatever the sampler returns, both unpadded edges come
|
|
# back as photo.
|
|
stitched = apply_stitch(stitcher, torch.rand(image.shape, generator=generator))
|
|
self.assertTrue(torch.equal(stitched[:, y0:y1, :1], image[:, y0:y1, :1]))
|
|
self.assertTrue(torch.equal(stitched[:, y0:y1, -1:], image[:, y0:y1, -1:]))
|
|
|
|
def test_validation_and_fingerprint(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
self.assertIs(cls.VALIDATE_INPUTS(image=path), True)
|
|
missing = str(Path(tmp) / "gone.png")
|
|
self.assertIn("Load Image + Pad", cls.VALIDATE_INPUTS(image=missing))
|
|
first = cls.IS_CHANGED(image=path, pad_left=1, feather=0)
|
|
second = cls.IS_CHANGED(image=path, pad_left=2, feather=0)
|
|
self.assertIsInstance(first, str)
|
|
self.assertNotEqual(first, second)
|
|
|
|
def test_source_image_is_an_appended_optional_input(self):
|
|
cls = self.make_node()
|
|
types = cls.INPUT_TYPES()
|
|
self.assertEqual(
|
|
list(types["required"]),
|
|
[
|
|
"image", "pad_left", "pad_top", "pad_right", "pad_bottom", "mode",
|
|
"fill_color", "backdrop_blur", "feather", "canvas_multiple",
|
|
"target_megapixels",
|
|
],
|
|
)
|
|
self.assertEqual(types["optional"]["source_image"][0], "IMAGE")
|
|
|
|
def test_a_wired_source_replaces_the_file(self):
|
|
import tempfile
|
|
|
|
cls = self.make_node()
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
(file_image, file_mask, *_), _ = self.run_node(cls, path)
|
|
# Same pixels as the file, plus an alpha channel a wire may carry;
|
|
# the file choice points nowhere, and must not be read.
|
|
rgb = torch.tensor([30, 180, 90], dtype=torch.float32) / 255.0
|
|
rgba = torch.cat([rgb.expand(1, 48, 64, 3), torch.zeros(1, 48, 64, 1)], dim=-1)
|
|
(image, mask, width, height, _, reference), _ = self.run_node(
|
|
cls, str(Path(tmp) / "gone.png"), source_image=rgba
|
|
)
|
|
self.assertEqual((width, height), (72, 56))
|
|
self.assertTrue(torch.allclose(image, file_image, atol=1e-6))
|
|
self.assertTrue(torch.equal(mask, file_mask))
|
|
self.assertEqual(int(reference.shape[-1]), 3)
|
|
|
|
def test_a_wired_batch_pads_every_frame_the_same_way(self):
|
|
cls = self.make_node()
|
|
batch = rand_image(3, 40, 30, seed=7)
|
|
(image, mask, width, height, _, _), _ = self.run_node(cls, "unused.png", source_image=batch)
|
|
self.assertEqual(tuple(image.shape), (3, height, width, 3))
|
|
for frame in range(3):
|
|
self.assertTrue(torch.equal(image[frame, 3:43, 2:32], batch[frame]))
|
|
|
|
def test_a_wired_run_hands_the_canvas_a_small_preview(self):
|
|
import tempfile
|
|
from unittest.mock import patch
|
|
|
|
import nodes.node_load_image_pad as module
|
|
|
|
cls = self.make_node()
|
|
source = rand_image(1, 900, 1200, seed=3)
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
fake = type("FolderPaths", (), {"get_temp_directory": staticmethod(lambda: tmp)})
|
|
with patch.object(module, "folder_paths", fake):
|
|
out = getattr(cls(), cls.FUNCTION)(
|
|
image="unused.png", pad_left=0, pad_top=0, pad_right=64, pad_bottom=0,
|
|
mode="color", fill_color="#808080", backdrop_blur=0.5, feather=0,
|
|
canvas_multiple=8, target_megapixels=0.0, source_image=source,
|
|
unique_id="12:5",
|
|
)
|
|
ui = out["ui"]
|
|
self.assertEqual(ui["ausboss_pad_source"], [[1200, 900]])
|
|
ref = ui["ausboss_pad_preview"][0]
|
|
self.assertEqual(ref, {"filename": "12_5.jpg", "subfolder": module.STAGE_PREVIEW_SUBFOLDER, "type": "temp"})
|
|
from PIL import Image
|
|
|
|
with Image.open(Path(tmp) / ref["subfolder"] / ref["filename"]) as preview:
|
|
self.assertLessEqual(max(preview.size), 512)
|
|
self.assertEqual(len(out["result"]), 6)
|
|
# A file-driven run keeps returning the plain tuple.
|
|
with tempfile.TemporaryDirectory() as tmp:
|
|
path = self.write_image(tmp)
|
|
plain, _ = self.run_node(cls, path)
|
|
self.assertIsInstance(plain, tuple)
|
|
|
|
def test_validation_skips_the_file_when_a_source_is_wired(self):
|
|
import inspect
|
|
|
|
cls = self.make_node()
|
|
# Named inputs only: **kwargs would switch off ComfyUI's own list and
|
|
# range checks for every input on the node.
|
|
self.assertIsNone(inspect.getfullargspec(cls.VALIDATE_INPUTS).varkw)
|
|
# ComfyUI passes a linked input to validation as None; an unwired
|
|
# optional input is left out.
|
|
self.assertIs(cls.VALIDATE_INPUTS(image="gone.png", source_image=None), True)
|
|
self.assertIn("Load Image + Pad", cls.VALIDATE_INPUTS(image="gone.png"))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|