81 lines
2.9 KiB
Python
81 lines
2.9 KiB
Python
"""Fractal interpolation — fill masked regions using fractal (self-similar) synthesis."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import numpy as np
|
|
|
|
from backend.node_registry import register_node
|
|
from backend.data_types import DataField
|
|
from backend.nodes.helpers import mask_to_bool
|
|
|
|
|
|
@register_node(display_name="Fractal Interpolation")
|
|
class FractalInterpolation:
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"field": ("DATA_FIELD",),
|
|
"mask": ("IMAGE",),
|
|
"iterations": ("INT", {"default": 200, "min": 10, "max": 5000, "step": 10}),
|
|
}
|
|
}
|
|
|
|
OUTPUTS = (
|
|
('DATA_FIELD', 'filled'),
|
|
)
|
|
FUNCTION = "process"
|
|
|
|
DESCRIPTION = (
|
|
"Fill masked regions using fractal interpolation. Matches the spectral "
|
|
"characteristics of the surrounding surface to produce natural-looking "
|
|
"infill that preserves texture. Better than Laplace for rough surfaces. "
|
|
)
|
|
|
|
KEYWORDS = ("inpaint", "fill", "hole", "infill")
|
|
|
|
def process(self, field: DataField, mask: np.ndarray, iterations: int) -> tuple:
|
|
data = np.asarray(field.data, dtype=np.float64).copy()
|
|
hole = mask_to_bool(mask)
|
|
|
|
if not hole.any():
|
|
return (field.replace(data=data),)
|
|
|
|
# Step 1: Estimate power spectrum from valid (unmasked) data
|
|
valid_data = data.copy()
|
|
valid_mean = data[~hole].mean() if (~hole).any() else 0.0
|
|
valid_data[hole] = valid_mean
|
|
|
|
fft_valid = np.fft.fft2(valid_data)
|
|
power = np.abs(fft_valid) ** 2
|
|
|
|
# Step 2: Generate fractal noise matching the power spectrum
|
|
rng = np.random.default_rng(42)
|
|
phases = rng.uniform(0, 2 * np.pi, data.shape)
|
|
noise_fft = np.sqrt(power) * np.exp(1j * phases)
|
|
noise = np.real(np.fft.ifft2(noise_fft))
|
|
|
|
# Normalize noise to match local statistics around masked region
|
|
if (~hole).any():
|
|
noise = (noise - noise[~hole].mean()) / max(noise[~hole].std(), 1e-30) * \
|
|
data[~hole].std() + data[~hole].mean()
|
|
|
|
# Step 3: Initialize masked pixels with fractal noise, then blend
|
|
# with Laplace relaxation for smooth boundaries
|
|
data[hole] = noise[hole]
|
|
|
|
# Relax boundaries to ensure continuity
|
|
padded = np.pad(data, 1, mode='edge')
|
|
hole_padded = np.pad(hole, 1, mode='constant', constant_values=False)
|
|
|
|
for _ in range(iterations):
|
|
avg = (padded[:-2, 1:-1] + padded[2:, 1:-1] +
|
|
padded[1:-1, :-2] + padded[1:-1, 2:]) / 4.0
|
|
# Blend: 90% fractal noise + 10% relaxation to smooth boundaries
|
|
blend = 0.1
|
|
new_vals = (1.0 - blend) * padded[1:-1, 1:-1][hole] + blend * avg[hole]
|
|
padded[1:-1, 1:-1][hole] = new_vals
|
|
|
|
data = padded[1:-1, 1:-1].copy()
|
|
return (field.replace(data=data),)
|