Ewald
831b072943
Core (nodes/core/): alpha, masks, batch consolidation, blend, transform matrix, bbox, parameters, validation engines shared across nodes, each with unit tests. Fix an unconditional RGB clamp to [0,1] that silently destroyed HDR/negative values in Merge, Transform, Color Correct, Comp, CornerPin, Distort, Spherize, Crop Stitch, and the shared masked-compositing helpers. Alpha stays clamped everywhere; RGB no longer does. Fix Grain applying `amount` twice in overlay/soft_light modes (quadratic instead of linear falloff). Add STMap mode to Distort (absolute UV remap vs relative displacement). Add bbox output/input wiring between Crop and Crop Stitch. Add edge_grow and edge_color_match to Crop Stitch. Add despill to Keyer. Add shuffle mode to Channel (R/G/B/A routed from any source). Add size/luma_response to Grain. Add grid mode to Constant. Add frame/layer accumulation limits to Preview and Draw, plus a font-path containment fix and supersampling in Text. Make Frame Range and Append recognize a real ComfyUI VIDEO input (not only this pack's internal media wrapper), preserving true fps/audio instead of silently dropping them, and add a native VIDEO output alongside the existing IMAGE output on both. Add node_replacements.py (empty registry, wired into the extension lifecycle) for future schema migrations. Add a CI matrix pinning the verified requires-comfyui floor, and .comfyignore for Registry packaging.
72 lines
1.9 KiB
Python
72 lines
1.9 KiB
Python
from __future__ import annotations
|
|
|
|
import torch
|
|
|
|
from nodes.core.alpha import premultiply, replace_alpha, split_rgba, unpremultiply
|
|
|
|
|
|
def test_split_rgba_with_alpha():
|
|
image = torch.rand((2, 4, 4, 4))
|
|
|
|
rgb, alpha = split_rgba(image)
|
|
|
|
assert rgb.shape == (2, 4, 4, 3)
|
|
assert alpha.shape == (2, 4, 4, 1)
|
|
assert torch.allclose(rgb, image[..., :3])
|
|
assert torch.allclose(alpha, image[..., 3:4])
|
|
|
|
|
|
def test_split_rgba_without_alpha():
|
|
image = torch.rand((2, 4, 4, 3))
|
|
|
|
rgb, alpha = split_rgba(image)
|
|
|
|
assert alpha is None
|
|
assert torch.allclose(rgb, image)
|
|
|
|
|
|
def test_premultiply_unpremultiply_round_trip():
|
|
rgb = torch.rand((1, 4, 4, 3)) * 2.0 - 0.5 # includes negative and >1 values
|
|
alpha = torch.rand((1, 4, 4, 1)).clamp(0.2, 1.0) # keep away from zero
|
|
|
|
premult = premultiply(rgb, alpha)
|
|
straight = unpremultiply(premult, alpha)
|
|
|
|
assert torch.allclose(straight, rgb, atol=1e-5)
|
|
|
|
|
|
def test_premultiply_none_alpha_is_noop():
|
|
rgb = torch.rand((1, 4, 4, 3))
|
|
|
|
assert torch.allclose(premultiply(rgb, None), rgb)
|
|
assert torch.allclose(unpremultiply(rgb, None), rgb)
|
|
|
|
|
|
def test_unpremultiply_zero_alpha_returns_zero_not_nan():
|
|
rgb = torch.full((1, 1, 1, 3), 5.0)
|
|
alpha = torch.zeros((1, 1, 1, 1))
|
|
|
|
result = unpremultiply(rgb, alpha)
|
|
|
|
assert torch.isfinite(result).all()
|
|
assert torch.allclose(result, torch.zeros_like(result))
|
|
|
|
|
|
def test_unpremultiply_preserves_hdr_values():
|
|
rgb = torch.tensor([[[[4.0, -1.0, 2.0]]]])
|
|
alpha = torch.ones((1, 1, 1, 1))
|
|
|
|
result = unpremultiply(rgb, alpha)
|
|
|
|
assert torch.allclose(result, rgb)
|
|
|
|
|
|
def test_replace_alpha_clamps_and_keeps_rgb():
|
|
image = torch.rand((1, 4, 4, 4)) * 2.0
|
|
new_alpha = torch.full((1, 4, 4, 1), 1.5)
|
|
|
|
out = replace_alpha(image, new_alpha)
|
|
|
|
assert torch.allclose(out[..., :3], image[..., :3])
|
|
assert torch.allclose(out[..., 3:4], torch.ones_like(new_alpha))
|