Claude
1567ea6de9
Critical fixes: - AnyAspectRatio: remove dead duplicate calculation that was overwriting itself with a wrong formula (correct formula kept on lines 55-56) - LoadImageResizer: fix trailing comma that made resized_mask a tuple instead of a value; properly convert alpha channel to float32 tensor - openAI_PoP: replace deprecated openai v0 API (openai.Image.create, openai.error.*) with modern openai>=1.0 client; fix hardcoded Windows backslash path with os.path.dirname(__file__); fix log/image dirs to be relative to module file instead of CWD - LoraStackLoaders: add missing `import comfy.sd` (was NameError at runtime); fix filter from l[0] (switch, never 'None') to l[1] (lora_name); fix `lora_name is None` to `== 'None'` for string comparison; fix display name mapping key LoraStackLoader10 -> LoraStackLoader10_PoP High severity fixes: - Conditioning: guard std() divisions with `if std > 0` to prevent NaN/Inf crash when tensor has zero variance - EfficientAttention: move dim_head calculation after dimension truncation so reshape is always valid; add divisibility check; fix output reshape to use min_dim not dim_q - VAEEncodeDecodeLoader: remove 5 debug print statements from decode() - CNutil: remove 3 debug print statements from resize_to_resolution() Minor fixes: - AdaptiveCannyDetector: fix `Category` -> `CATEGORY` (case-sensitive, ComfyUI was ignoring the node category) - LoadImageResizer: remove duplicate CATEGORY = "image" definition - requirements.txt: remove unused matplotlib/seaborn; add missing Pillow https://claude.ai/code/session_01QPLKoy7P41H3QPB6tMrpPh
37 lines
1.2 KiB
Python
37 lines
1.2 KiB
Python
import numpy as np
|
|
import cv2
|
|
from PIL import Image
|
|
|
|
|
|
|
|
def convert_to_3_channels(x):
|
|
if x.dtype == np.float32:
|
|
assert x.min() >= 0.0 and x.max() <= 1.0
|
|
x = (255.0 * x).astype(np.uint8)
|
|
elif x.dtype != np.uint8:
|
|
raise ValueError("Unsupported dtype")
|
|
x = np.squeeze(x, axis=0)
|
|
if x.ndim == 2:
|
|
x = x[:, :, None]
|
|
assert x.ndim == 3
|
|
H, W, C = x.shape
|
|
assert C in (1, 3, 4)
|
|
if C == 3:
|
|
return x
|
|
if C == 1:
|
|
return np.repeat(x, 3, axis=2)
|
|
if C == 4:
|
|
color = x[:, :, :3].astype(np.float32)
|
|
alpha = x[:, :, 3:4].astype(np.float32) / 255.0
|
|
y = color * alpha + 255.0 * (1.0 - alpha)
|
|
return y.clip(0, 255).astype(np.uint8)
|
|
|
|
# Resize an image to a resolution, preserving aspect ratio
|
|
def resize_to_resolution(input_image, resolution):
|
|
H, W, C = input_image.shape
|
|
k = resolution / min(H, W)
|
|
H_new = int(np.round(H * k / 64.0)) * 64
|
|
W_new = int(np.round(W * k / 64.0)) * 64
|
|
interpolation = cv2.INTER_AREA if k < 1.0 else cv2.INTER_LINEAR
|
|
resized_image = cv2.resize(input_image, (W_new, H_new), interpolation=interpolation)
|
|
return resized_image |