MONKEYFOREVER2andClaude Fable 5
5bb0c40118
- Rework nodes.py: noise-adaptive smart_auto (h = 1.15*sigma heuristic), edge-aware detail recovery, pure-numpy guided filter, LAB bilateral, wavelet BayesShrink, TV, bm3d with NLM fallback, optional sharpening - Rewrite js/denoiser_widgets.js: themed panel with segmented method picker, contextual sliders, collapsible advanced section - Add test_node.py standalone smoke test (all methods + identity check) - Update README for current methods, UI, and testing; add .gitignore; mark bm3d optional in requirements.txt Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
57 lines
2.2 KiB
Python
57 lines
2.2 KiB
Python
"""Standalone smoke test — run with the ComfyUI venv python."""
|
|
import numpy as np
|
|
import torch
|
|
import cv2
|
|
import nodes
|
|
|
|
|
|
def make_test_batch():
|
|
"""Two synthetic images: edges+gradient with Gaussian noise, and one clean."""
|
|
h, w = 128, 160
|
|
img = np.zeros((h, w, 3), np.float32)
|
|
img[:, :, 0] = np.linspace(0, 1, w)[None, :] # gradient
|
|
cv2.rectangle(img, (30, 30), (90, 90), (0.2, 0.9, 0.4), -1) # hard edges
|
|
cv2.circle(img, (120, 64), 25, (0.9, 0.3, 0.2), -1)
|
|
noisy = np.clip(img + np.random.normal(0, 0.05, img.shape), 0, 1)
|
|
batch = np.stack([noisy, img]).astype(np.float32)
|
|
return torch.from_numpy(batch)
|
|
|
|
|
|
def main():
|
|
node = nodes.AdvancedImageDenoiser()
|
|
batch = make_test_batch()
|
|
print(f"skimage={nodes.HAS_SKIMAGE} bm3d={nodes.HAS_BM3D}")
|
|
|
|
for method in nodes.AdvancedImageDenoiser.METHODS:
|
|
out, report = node.denoise(batch, method, strength=0.15,
|
|
detail_recovery=0.35)
|
|
assert out.shape == batch.shape, f"{method}: shape {out.shape}"
|
|
assert out.dtype == torch.float32
|
|
assert 0.0 <= out.min() and out.max() <= 1.0
|
|
# Denoising should reduce noise: compare residual vs the clean frame
|
|
noisy_mse = ((batch[0] - batch[1]) ** 2).mean().item()
|
|
out_mse = ((out[0] - batch[1]) ** 2).mean().item()
|
|
print(f"{method:18s} MSE noisy={noisy_mse:.5f} -> out={out_mse:.5f} "
|
|
f"| {report.splitlines()[0]}")
|
|
|
|
# Sharpening + blend paths
|
|
out, _ = node.denoise(batch, "smart_auto", 0.15, 0.35,
|
|
blend_original=0.2, sharpen_mode="luminance_only",
|
|
sharpen_amount=0.3)
|
|
assert out.shape == batch.shape
|
|
out, _ = node.denoise(batch, "smart_auto", 0.15, 0.35,
|
|
sharpen_mode="unsharp_mask", sharpen_amount=0.3)
|
|
assert out.shape == batch.shape
|
|
|
|
# Zero-strength should be near-identity
|
|
out, _ = node.denoise(batch, "non_local_means", 0.0, 0.0,
|
|
luminance_strength=0.0, chroma_strength=0.0)
|
|
diff = (out - batch).abs().max().item()
|
|
assert diff < 0.01, f"zero strength changed image by {diff}"
|
|
print("zero-strength identity OK")
|
|
print("ALL TESTS PASSED")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|