anima-style-factor-intervention-v4 / code /factor_interventions.py
ij's picture
Add files using upload-large-folder tool
02443ff verified
Raw
History Blame Contribute Delete
13.2 kB
from __future__ import annotations
import math
import cv2
import numpy as np
from PIL import Image
from .image_preprocess import canonical_square, decode_rgb
FACTORS = ("line", "color", "texture", "geometry")
INTERVENTION_VERSION = "lens-safe-v4"
TRAIN_FAMILIES = {
"line": ("dilate_erode", "blur_sharpen", "darkness_contrast"),
"color": ("palette_remap", "split_tone", "tone_curve"),
"texture": ("smooth_detail", "frequency", "grain_noise"),
"geometry": ("crop_zoom_translate", "perspective", "lens_warp"),
}
VALIDATION_FAMILIES = {
"line": "edge_overlay",
"color": "channel_mixer",
"texture": "median_speckle",
"geometry": "shear",
}
LEVELS = {"weak": 0.50, "medium": 0.80, "strong": 1.0}
def _uint8(array: np.ndarray) -> np.ndarray:
return np.clip(array, 0, 255).astype(np.uint8)
def _edges(array: np.ndarray) -> np.ndarray:
gray = cv2.cvtColor(array, cv2.COLOR_RGB2GRAY)
return cv2.Canny(gray, 60, 150).astype(np.float32) / 255.0
def _pixel_transform(
image: Image.Image, factor: str, family: str, signed: float, seed: int
) -> Image.Image:
magnitude = abs(signed)
direction = 1 if signed >= 0 else -1
array = np.asarray(image.convert("RGB"), dtype=np.uint8)
if factor == "line":
if family in {"dilate_erode", "edge_overlay"}:
edge = _edges(array)
kernel_size = 3 if magnitude <= 0.5 else 5 if magnitude < 1.0 else 7
kernel = np.ones((kernel_size, kernel_size), np.uint8)
edge = cv2.dilate(edge, kernel)
if direction > 0:
alpha = (0.72 if family == "edge_overlay" else 0.64) * magnitude
result = array.astype(np.float32) * (1.0 - alpha * edge[..., None])
else:
smooth = cv2.bilateralFilter(array, 11, 80, 9).astype(np.float32)
weight = np.clip(0.90 * magnitude * edge, 0.0, 0.90)[..., None]
result = array * (1.0 - weight) + smooth * weight
elif family == "blur_sharpen":
blurred = cv2.GaussianBlur(array, (0, 0), 1.2 + 2.4 * magnitude)
result = blurred if direction < 0 else cv2.addWeighted(
array, 1.0 + 1.45 * magnitude, blurred, -1.45 * magnitude, 0
)
elif family == "darkness_contrast":
edge = cv2.GaussianBlur(_edges(array), (0, 0), 1.0)
if direction > 0:
result = array.astype(np.float32) * (1.0 - 0.70 * magnitude * edge[..., None])
else:
smooth = cv2.GaussianBlur(array, (0, 0), 2.6)
weight = np.clip(0.90 * magnitude * edge, 0.0, 0.90)[..., None]
result = array * (1.0 - weight) + smooth * weight
else:
raise ValueError(f"unknown line family: {family}")
return Image.fromarray(_uint8(result))
if factor == "color":
unit = array.astype(np.float32) / 255.0
if family == "palette_remap":
hsv = cv2.cvtColor(array, cv2.COLOR_RGB2HSV).astype(np.float32)
hsv[..., 0] = np.mod(hsv[..., 0] + direction * 30.0 * magnitude, 180.0)
saturation = 1.0 + 0.95 * magnitude if direction > 0 else 1.0 - 0.70 * magnitude
hsv[..., 1] *= saturation
hsv[..., 2] = 255.0 * np.power(
hsv[..., 2] / 255.0, math.exp(-direction * 0.38 * magnitude)
)
result = cv2.cvtColor(_uint8(hsv), cv2.COLOR_HSV2RGB).astype(np.float32)
balance = direction * 0.16 * magnitude
result[..., 0] *= 1.0 + balance
result[..., 2] *= 1.0 - balance
elif family == "split_tone":
luminance = np.sum(unit * np.array([0.213, 0.715, 0.072], np.float32), axis=2)
shadows = np.power(1.0 - luminance, 1.4)[..., None]
highlights = np.power(luminance, 1.4)[..., None]
cool_shadow = np.array([-0.16, 0.02, 0.28], np.float32)
warm_highlight = np.array([0.30, 0.12, -0.14], np.float32)
result = unit + direction * magnitude * (
shadows * cool_shadow + highlights * warm_highlight
)
result = 0.5 + (result - 0.5) * (1.0 + 0.38 * magnitude)
result = _uint8(result * 255.0)
hsv = cv2.cvtColor(result, cv2.COLOR_RGB2HSV).astype(np.float32)
hsv[..., 0] = np.mod(hsv[..., 0] + direction * 12.0 * magnitude, 180.0)
hsv[..., 1] *= 1.0 + 0.40 * magnitude
result = cv2.cvtColor(_uint8(hsv), cv2.COLOR_HSV2RGB)
elif family == "tone_curve":
channel_gamma = np.exp(
direction * magnitude * np.array([-0.62, -0.18, 0.48], np.float32)
)
result = np.power(np.clip(unit, 0.0, 1.0), channel_gamma)
contrast = 1.0 + direction * 0.48 * magnitude
result = 0.5 + (result - 0.5) * contrast
luminance = np.sum(result * np.array([0.213, 0.715, 0.072], np.float32), axis=2, keepdims=True)
result = luminance + (result - luminance) * (1.0 + 0.55 * magnitude)
result *= 1.0 + direction * 0.10 * magnitude
result = result * 255.0
elif family == "channel_mixer":
delta = np.array(
[[0.22, 0.12, -0.20], [-0.12, 0.20, 0.08], [0.08, -0.20, 0.24]],
np.float32,
)
matrix = np.eye(3, dtype=np.float32) + direction * magnitude * delta
result = unit @ matrix.T
result = np.power(
np.clip(result, 0.0, 1.0), math.exp(-direction * 0.45 * magnitude)
)
result = 255.0 * result
else:
raise ValueError(f"unknown color family: {family}")
return Image.fromarray(_uint8(result))
if factor == "texture":
result = array.astype(np.float32)
if family == "smooth_detail":
smooth = cv2.bilateralFilter(array, 13, 70 + 55 * magnitude, 11)
if direction < 0:
result = cv2.addWeighted(array, 1.0 - 0.85 * magnitude, smooth, 0.85 * magnitude, 0)
else:
result = array + (array.astype(np.float32) - smooth) * 1.35 * magnitude
elif family == "frequency":
low = cv2.GaussianBlur(array, (0, 0), 1.5 + 1.5 * magnitude)
high = array.astype(np.float32) - low.astype(np.float32)
result = array + direction * high * 1.55 * magnitude
elif family in {"grain_noise", "median_speckle"}:
if direction < 0:
kernel = 5 if magnitude >= 0.8 else 3
median = cv2.medianBlur(array, kernel)
result = cv2.addWeighted(array, 1.0 - 0.85 * magnitude, median, 0.85 * magnitude, 0)
else:
rng = np.random.default_rng(seed)
noise = rng.normal(0, 26.0 * magnitude, array.shape[:2]).astype(np.float32)
noise = cv2.GaussianBlur(noise, (0, 0), 0.35)[..., None]
result = array.astype(np.float32) + noise
else:
raise ValueError(f"unknown texture family: {family}")
return Image.fromarray(_uint8(result))
raise ValueError(f"pixel transform does not support factor: {factor}")
def _apply_homography(
image: Image.Image, matrix: np.ndarray
) -> tuple[Image.Image, callable]:
array = np.asarray(image)
height, width = array.shape[:2]
warped = cv2.warpPerspective(
array, matrix, (width, height), flags=cv2.INTER_LANCZOS4, borderMode=cv2.BORDER_REFLECT_101
)
def map_points(points: np.ndarray) -> np.ndarray:
return cv2.perspectiveTransform(points.astype(np.float32)[None], matrix)[0]
return Image.fromarray(warped), map_points
def _geometry_transform(
image: Image.Image, family: str, signed: float, seed: int
) -> tuple[Image.Image, callable]:
magnitude = abs(signed)
direction = 1 if signed >= 0 else -1
width, height = image.size
center = np.array([width / 2, height / 2], dtype=np.float32)
if family == "crop_zoom_translate":
rng = np.random.default_rng(seed)
scale = 1.0 + direction * 0.20 * magnitude
shift = np.array(
[direction * 0.10 * width, (1 if rng.integers(2) else -1) * 0.07 * height],
dtype=np.float32,
) * magnitude
matrix = np.array(
[[scale, 0, center[0] * (1 - scale) + shift[0]], [0, scale, center[1] * (1 - scale) + shift[1]], [0, 0, 1]],
dtype=np.float32,
)
return _apply_homography(image, matrix)
if family in {"perspective", "shear"}:
source = np.array([[0, 0], [width - 1, 0], [width - 1, height - 1], [0, height - 1]], np.float32)
if family == "perspective":
delta = direction * 0.15 * width * magnitude
target = source + np.array([[delta, 0], [-delta, delta * 0.35], [delta, 0], [-delta, -delta * 0.35]], np.float32)
else:
delta = direction * 0.18 * width * magnitude
target = source + np.array([[delta, 0], [delta, 0], [-delta, 0], [-delta, 0]], np.float32)
return _apply_homography(image, cv2.getPerspectiveTransform(source, target))
if family == "lens_warp":
# Keep radial deformation perceptible without creating a bubble silhouette.
radial_strength = float(
np.interp(magnitude, [0.5, 0.8, 1.0], [0.035, 0.055, 0.075])
)
k = direction * radial_strength
yy, xx = np.indices((height, width), dtype=np.float32)
xd = (xx - center[0]) / (width / 2)
yd = (yy - center[1]) / (height / 2)
xs, ys = xd.copy(), yd.copy()
for _ in range(4):
radius2 = xs * xs + ys * ys
factor = 1.0 + k * radius2
xs, ys = xd / factor, yd / factor
map_x = xs * (width / 2) + center[0]
map_y = ys * (height / 2) + center[1]
warped = cv2.remap(
np.asarray(image), map_x, map_y, cv2.INTER_LANCZOS4, borderMode=cv2.BORDER_REFLECT_101
)
def map_points(points: np.ndarray) -> np.ndarray:
normalized = (points - center) / np.array([width / 2, height / 2])
radius2 = np.square(normalized).sum(axis=1, keepdims=True)
return center + normalized * (1.0 + k * radius2) * np.array([width / 2, height / 2])
return Image.fromarray(warped), map_points
raise ValueError(f"unknown geometry family: {family}")
def _face_box_in_square(metadata: dict, crop_box: tuple[int, int, int, int], size: int) -> np.ndarray | None:
box = (metadata.get("face_detection") or {}).get("primary_box")
if box is None:
return None
left, top, right, bottom = crop_box
scale_x = size / (right - left)
scale_y = size / (bottom - top)
x0, y0, x1, y1 = map(float, box)
return np.array(
[[(x0 - left) * scale_x, (y0 - top) * scale_y], [(x1 - left) * scale_x, (y0 - top) * scale_y], [(x1 - left) * scale_x, (y1 - top) * scale_y], [(x0 - left) * scale_x, (y1 - top) * scale_y]],
dtype=np.float32,
)
def _crop_face(image: Image.Image, points: np.ndarray, size: int, padding: float = 0.25) -> Image.Image:
x0, y0 = points.min(axis=0)
x1, y1 = points.max(axis=0)
if x1 <= 0 or y1 <= 0 or x0 >= image.width or y0 >= image.height:
raise ValueError("transformed face left the image")
side = max(x1 - x0, y1 - y0) * (1 + 2 * padding)
if side < 8:
raise ValueError("transformed face is too small")
cx, cy = (x0 + x1) / 2, (y0 + y1) / 2
box = (max(0, cx - side / 2), max(0, cy - side / 2), min(image.width, cx + side / 2), min(image.height, cy + side / 2))
crop = image.crop(tuple(map(int, map(round, box))))
if min(crop.size) < 4:
raise ValueError("invalid transformed face crop")
return crop.resize((size, size), Image.Resampling.LANCZOS)
def apply_intervention(
full_bytes: bytes,
face_bytes: bytes | None,
metadata: dict,
spec: dict,
*,
size: int = 512,
) -> tuple[Image.Image, Image.Image | None]:
full, crop_box = canonical_square(decode_rgb(full_bytes), size)
signed = float(spec["signed_intensity"])
seed = int(spec["operation_seed"])
factor = str(spec["factor"])
family = str(spec["family"])
if factor == "geometry":
transformed, map_points = _geometry_transform(full, family, signed, seed)
points = _face_box_in_square(metadata, crop_box, size)
face = None
if points is not None:
try:
face = _crop_face(transformed, map_points(points), size)
except ValueError:
face = None
if face is None and face_bytes is not None:
face_source, _ = canonical_square(decode_rgb(face_bytes), size)
face, _ = _geometry_transform(face_source, family, signed, seed ^ 0x5A17)
return transformed, face
transformed = _pixel_transform(full, factor, family, signed, seed)
face = None
if face_bytes is not None:
face, _ = canonical_square(decode_rgb(face_bytes), size)
face = _pixel_transform(face, factor, family, signed, seed ^ 0x5A17)
return transformed, face