ausboss
dd118c7386
- Optional resize of the transformed output to a pixel budget, with core Scale Image to Total Pixels semantics: megapixels x 1024^2, aspect preserved, each dimension rounded to resolution_steps. The image uses the chosen filter (lanczos/area/bicubic/bilinear/nearest-exact); the mask always resizes bilinear so feathered edges cannot ring. The math is mirrored py+js (scale_to_megapixels / scaleToMegapixels) and fixture-tested on both sides; the new widgets append after the stable V1 set so saved workflows keep loading, and IS_CHANGED fingerprints them. - A quick row under the node's canvas: Reset (rotation, crop, padding), a Feather toggle that remembers the amount it turns off, and the Resize toggle with its megapixel scrub box. - The editor gains a Resize output section (megapixels and steps as scrub controls, method select) and its status names the exact resized size. - The output size readout moved off the pixels: centered below the image, flipping above when clipped, and showing the resize target. - CHANGELOG entries for this round and the two commits before it.
324 lines
14 KiB
Python
324 lines
14 KiB
Python
from __future__ import annotations
|
|
|
|
import os
|
|
from pathlib import Path
|
|
import sys
|
|
import tempfile
|
|
import time
|
|
import unittest
|
|
import unittest.mock
|
|
|
|
import av
|
|
import numpy as np
|
|
from PIL import Image
|
|
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"]
|
|
|
|
from nodes._media_helpers import decode_video_frame, video_metadata
|
|
from nodes._transform_engine import (
|
|
scale_to_megapixels,
|
|
TransformSpec,
|
|
stable_file_fingerprint,
|
|
transform_pil,
|
|
transform_tensor_batch,
|
|
)
|
|
|
|
|
|
def solid(width=12, height=8, color=(30, 80, 140, 255)):
|
|
return Image.new("RGBA", (width, height), color)
|
|
|
|
|
|
class TransformEngineTests(unittest.TestCase):
|
|
def test_identity_preserves_dimensions_and_pixels(self):
|
|
source = solid()
|
|
output, mask, geometry = transform_pil(source, TransformSpec())
|
|
self.assertEqual(output.size, source.size)
|
|
self.assertEqual(geometry.output_width, 12)
|
|
self.assertEqual(np.asarray(mask).max(), 0)
|
|
self.assertTrue(np.all(np.asarray(output) == np.array([30, 80, 140])))
|
|
|
|
def test_source_alpha_is_part_of_mask(self):
|
|
source = solid(5, 5)
|
|
source.putpixel((2, 2), (255, 0, 0, 0))
|
|
_, mask, _ = transform_pil(source, TransformSpec())
|
|
self.assertEqual(mask.getpixel((2, 2)), 255)
|
|
self.assertEqual(mask.getpixel((0, 0)), 0)
|
|
|
|
def test_positive_and_negative_rotation_boundaries(self):
|
|
for angle in (-90, 90, 179.9, -179.9):
|
|
with self.subTest(angle=angle):
|
|
output, mask, geometry = transform_pil(solid(11, 7), TransformSpec(rotation_degrees=angle))
|
|
self.assertEqual(output.size, mask.size)
|
|
self.assertGreaterEqual(geometry.rotated_width, 7)
|
|
self.assertGreaterEqual(geometry.rotated_height, 7)
|
|
if abs(angle) % 90 > 0.01:
|
|
self.assertGreater(np.asarray(mask).max(), 0)
|
|
|
|
def test_crop_clamps_to_rotated_source(self):
|
|
output, _, geometry = transform_pil(
|
|
solid(20, 10), TransformSpec(crop_x=18, crop_y=9, crop_width=999, crop_height=999)
|
|
)
|
|
self.assertEqual(output.size, (2, 1))
|
|
self.assertEqual((geometry.crop_x, geometry.crop_y), (18, 9))
|
|
|
|
def test_ratio_is_enforced_in_backend(self):
|
|
output, _, _ = transform_pil(
|
|
solid(100, 100), TransformSpec(crop_aspect_ratio="16:9")
|
|
)
|
|
self.assertEqual(output.size, (100, 56))
|
|
|
|
def test_every_padding_side_changes_output_and_mask(self):
|
|
cases = {
|
|
"pad_left": ((7, 4), (0, 1)),
|
|
"pad_top": ((5, 6), (1, 0)),
|
|
"pad_right": ((7, 4), (6, 1)),
|
|
"pad_bottom": ((5, 6), (1, 5)),
|
|
}
|
|
for field, (expected_size, generated_pixel) in cases.items():
|
|
with self.subTest(field=field):
|
|
output, mask, _ = transform_pil(solid(5, 4), TransformSpec(**{field: 2}))
|
|
self.assertEqual(output.size, expected_size)
|
|
self.assertEqual(mask.getpixel(generated_pixel), 255)
|
|
|
|
def test_fill_color_is_used_for_padding_and_rotation_voids(self):
|
|
output, _, _ = transform_pil(
|
|
solid(4, 3, (255, 0, 0, 255)),
|
|
TransformSpec(rotation_degrees=45, pad_left=2, fill_color="#123456"),
|
|
)
|
|
self.assertEqual(output.getpixel((0, 0)), (0x12, 0x34, 0x56))
|
|
|
|
def test_feather_keeps_generated_pixels_and_softens_boundary(self):
|
|
_, mask, _ = transform_pil(solid(20, 20), TransformSpec(pad_left=10, feather=3))
|
|
values = np.asarray(mask)
|
|
self.assertEqual(values[:, :5].min(), 255)
|
|
self.assertTrue(np.any((values[:, 10:15] > 0) & (values[:, 10:15] < 255)))
|
|
# No seam: the ramp is still near full strength on the first kept
|
|
# column (the old max-with-blur approach dropped to ~127 here).
|
|
self.assertGreaterEqual(int(values[10, 10]), 200)
|
|
# And it decays moving further into kept content.
|
|
self.assertLess(int(values[10, 16]), int(values[10, 11]))
|
|
|
|
def test_feather_fades_image_into_fill(self):
|
|
output, _, _ = transform_pil(
|
|
solid(20, 20, (255, 255, 255, 255)),
|
|
TransformSpec(pad_left=10, feather=3, fill_color="#000000"),
|
|
)
|
|
pixels = np.asarray(output)
|
|
row = pixels[10]
|
|
self.assertLessEqual(int(row[9].max()), 40) # padding stays fill color
|
|
self.assertTrue(40 < int(row[11][0]) < 240) # visible blend band
|
|
self.assertGreaterEqual(int(row[25].min()), 250) # interior untouched
|
|
|
|
def test_canvas_multiple_rounds_only_right_and_bottom(self):
|
|
output, mask, geometry = transform_pil(
|
|
solid(101, 99), TransformSpec(canvas_multiple=8)
|
|
)
|
|
self.assertEqual(output.size, (104, 104))
|
|
self.assertEqual((geometry.pad_left, geometry.pad_top), (0, 0))
|
|
self.assertEqual((geometry.pad_right, geometry.pad_bottom), (3, 5))
|
|
self.assertEqual(mask.getpixel((103, 103)), 255)
|
|
|
|
def test_tensor_batch_preserves_bhwc_and_bhw(self):
|
|
source = torch.zeros((3, 9, 7, 3), dtype=torch.float32)
|
|
output, mask, _ = transform_tensor_batch(source, TransformSpec(pad_bottom=1))
|
|
self.assertEqual(tuple(output.shape), (3, 10, 7, 3))
|
|
self.assertEqual(tuple(mask.shape), (3, 10, 7))
|
|
self.assertEqual(output.dtype, source.dtype)
|
|
|
|
def test_invalid_shape_and_oversized_padding_are_actionable(self):
|
|
with self.assertRaisesRegex(ValueError, "expected BHWC"):
|
|
transform_tensor_batch(torch.zeros((4, 4, 3)), TransformSpec())
|
|
with self.assertRaisesRegex(ValueError, "pad_left"):
|
|
transform_pil(solid(), TransformSpec(pad_left=40000))
|
|
|
|
def test_fingerprint_changes_with_inputs_and_file_state(self):
|
|
with tempfile.TemporaryDirectory() as directory:
|
|
path = Path(directory, "source.bin")
|
|
path.write_bytes(b"one")
|
|
first = stable_file_fingerprint(path, {"crop": 0})
|
|
second = stable_file_fingerprint(path, {"crop": 1})
|
|
self.assertNotEqual(first, second)
|
|
time.sleep(0.01)
|
|
path.write_bytes(b"two-two")
|
|
os.utime(path, None)
|
|
self.assertNotEqual(first, stable_file_fingerprint(path, {"crop": 0}))
|
|
|
|
|
|
class VideoDecodeTests(unittest.TestCase):
|
|
def setUp(self):
|
|
self._tempdir = tempfile.TemporaryDirectory()
|
|
self.directory = Path(self._tempdir.name)
|
|
|
|
def tearDown(self):
|
|
# Persistent scrub sessions hold the file open (locked on Windows);
|
|
# release them before the temporary directory is removed.
|
|
from nodes import _media_helpers
|
|
|
|
_media_helpers.close_scrub_sessions()
|
|
self._tempdir.cleanup()
|
|
|
|
def _write_video(self, path: Path, frames: int = 5, step: int = 45):
|
|
container = av.open(str(path), mode="w")
|
|
stream = container.add_stream("mpeg4", rate=5)
|
|
stream.width = 16
|
|
stream.height = 16
|
|
stream.pix_fmt = "yuv420p"
|
|
for index in range(frames):
|
|
array = np.zeros((16, 16, 3), dtype=np.uint8)
|
|
array[..., 0] = (index * step) % 256
|
|
frame = av.VideoFrame.from_ndarray(array, format="rgb24")
|
|
for packet in stream.encode(frame):
|
|
container.mux(packet)
|
|
for packet in stream.encode():
|
|
container.mux(packet)
|
|
container.close()
|
|
|
|
def test_first_middle_last_frame_decoding(self):
|
|
path = Path(self.directory, "frames.mp4")
|
|
self._write_video(path)
|
|
metadata = video_metadata(path)
|
|
self.assertEqual(metadata["width"], 16)
|
|
self.assertGreaterEqual(metadata["frame_count"], 5)
|
|
reds = []
|
|
for index in (0, 2, 4):
|
|
frame, actual_index, _ = decode_video_frame(path, "frame index", index, 0.0)
|
|
self.assertEqual(actual_index, index)
|
|
reds.append(float(np.asarray(frame)[..., 0].mean()))
|
|
self.assertLess(reds[0], reds[1])
|
|
self.assertLess(reds[1], reds[2])
|
|
|
|
def test_keyframe_seek_agrees_with_sequential_scan(self):
|
|
# 64 frames at GOP defaults spans several keyframes, so mid and late
|
|
# targets exercise the container.seek fast path against ground truth.
|
|
from nodes._media_helpers import _decode_sequential
|
|
|
|
path = Path(self.directory, "long.mp4")
|
|
self._write_video(path, frames=64, step=4)
|
|
fps = float(video_metadata(path)["fps"] or 0.0)
|
|
self.assertGreater(fps, 0)
|
|
for target in (0, 7, 33, 63):
|
|
with self.subTest(target=target):
|
|
fast_frame, fast_index, fast_time = decode_video_frame(
|
|
path, "frame index", target, 0.0
|
|
)
|
|
slow_frame, slow_index, slow_time = _decode_sequential(path, target, fps)
|
|
self.assertEqual(fast_index, slow_index)
|
|
self.assertAlmostEqual(fast_time, slow_time, places=4)
|
|
self.assertTrue(
|
|
np.array_equal(np.asarray(fast_frame), np.asarray(slow_frame))
|
|
)
|
|
|
|
def test_time_seek_mode_lands_on_matching_frame(self):
|
|
path = Path(self.directory, "timed.mp4")
|
|
self._write_video(path, frames=25, step=10)
|
|
# 5 fps: 2.0 seconds is exactly frame 10.
|
|
_, actual_index, actual_time = decode_video_frame(path, "time seconds", 0, 2.0)
|
|
self.assertEqual(actual_index, 10)
|
|
self.assertAlmostEqual(actual_time, 2.0, places=3)
|
|
|
|
def test_out_of_range_index_clamps_to_last_frame(self):
|
|
path = Path(self.directory, "short.mp4")
|
|
self._write_video(path, frames=6, step=40)
|
|
_, actual_index, _ = decode_video_frame(path, "frame index", 999999, 0.0)
|
|
self.assertEqual(actual_index, 5)
|
|
|
|
def test_session_forward_and_backward_decode_match_fresh_decode(self):
|
|
from nodes import _media_helpers
|
|
|
|
path = Path(self.directory, "session.mp4")
|
|
self._write_video(path, frames=40, step=6)
|
|
fps = float(video_metadata(path)["fps"] or 0.0)
|
|
decode_video_frame(path, "frame index", 5, 0.0)
|
|
# Forward within the gap uses the persistent session's decoder
|
|
# state; backward forces a re-seek on the same open container.
|
|
for target in (9, 2):
|
|
with self.subTest(target=target):
|
|
fast = decode_video_frame(path, "frame index", target, 0.0)
|
|
fresh = _media_helpers._decode_sequential(path, target, fps)
|
|
self.assertEqual(fast[1], fresh[1])
|
|
self.assertTrue(np.array_equal(np.asarray(fast[0]), np.asarray(fresh[0])))
|
|
|
|
def test_storyboard_builds_ready_payload(self):
|
|
import base64
|
|
from io import BytesIO
|
|
|
|
from nodes import _media_helpers
|
|
|
|
path = Path(self.directory, "board.mp4")
|
|
self._write_video(path, frames=30, step=8)
|
|
key = _media_helpers._file_key(path)
|
|
# Seed "building" so storyboard_payload does not spawn its
|
|
# background thread; the build below is fully deterministic.
|
|
with _media_helpers._STORYBOARDS_LOCK:
|
|
_media_helpers._STORYBOARDS[key] = {"status": "building"}
|
|
self.assertEqual(_media_helpers.storyboard_payload(path)["status"], "building")
|
|
_media_helpers._build_storyboard(path, key)
|
|
ready = _media_helpers.storyboard_payload(path)
|
|
self.assertEqual(ready["status"], "ready")
|
|
self.assertGreaterEqual(ready["count"], 1)
|
|
self.assertEqual(ready["times"], sorted(ready["times"]))
|
|
sprite_bytes = base64.b64decode(ready["sprite"].split(",", 1)[1])
|
|
with Image.open(BytesIO(sprite_bytes)) as sprite:
|
|
self.assertEqual(
|
|
sprite.size, (ready["tile_width"] * ready["count"], ready["tile_height"])
|
|
)
|
|
|
|
def test_metadata_cache_hits_and_invalidates_on_file_change(self):
|
|
from nodes import _media_helpers
|
|
|
|
path = Path(self.directory, "cached.mp4")
|
|
self._write_video(path, frames=4)
|
|
first = _media_helpers.cached_video_metadata(path)
|
|
with unittest.mock.patch.object(
|
|
_media_helpers, "video_metadata", side_effect=AssertionError("cache miss")
|
|
):
|
|
# Same file state: served from cache, video_metadata untouched.
|
|
self.assertEqual(_media_helpers.cached_video_metadata(path), first)
|
|
self._write_video(path, frames=8)
|
|
os.utime(path, None)
|
|
refreshed = _media_helpers.cached_video_metadata(path)
|
|
self.assertGreaterEqual(int(refreshed["frame_count"]), 8)
|
|
|
|
|
|
class LocalPreviewGateTests(unittest.TestCase):
|
|
def test_disabled_by_default_outside_managed_folders(self):
|
|
from nodes._media_helpers import local_preview_allowed
|
|
|
|
with unittest.mock.patch.dict(os.environ, {}, clear=False):
|
|
os.environ.pop("AUSBOSS_TRANSFORM_LOCAL_PREVIEW", None)
|
|
self.assertFalse(local_preview_allowed(str(Path.home() / "video.mp4")))
|
|
self.assertFalse(local_preview_allowed(""))
|
|
|
|
def test_environment_flag_enables_previews(self):
|
|
from nodes._media_helpers import local_preview_allowed
|
|
|
|
with unittest.mock.patch.dict(
|
|
os.environ, {"AUSBOSS_TRANSFORM_LOCAL_PREVIEW": "1"}
|
|
):
|
|
self.assertTrue(local_preview_allowed(str(Path.home() / "video.mp4")))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|
|
|
|
|
|
class TestScaleToMegapixels(unittest.TestCase):
|
|
def test_square_megapixel_is_1024(self):
|
|
self.assertEqual(scale_to_megapixels(1024, 1024, 1.0, 1), (1024, 1024))
|
|
self.assertEqual(scale_to_megapixels(512, 512, 4.0, 1), (2048, 2048))
|
|
|
|
def test_hd_at_one_megapixel_on_64_grid(self):
|
|
# The everyday case: 1920x1080 to ~1MP on a VAE-friendly grid.
|
|
self.assertEqual(scale_to_megapixels(1920, 1080, 1.0, 64), (1344, 768))
|
|
|
|
def test_steps_round_each_dimension_independently(self):
|
|
self.assertEqual(scale_to_megapixels(1000, 707, 1.0, 8), (1216, 864))
|
|
|
|
def test_never_below_one_step(self):
|
|
width, height = scale_to_megapixels(100, 100, 0.01, 64)
|
|
self.assertEqual((width, height), (128, 128))
|