In generative AI (e.g., Stable Diffusion), is a critical slider ranging from 0.0 to 1.0 .
AI denoisers use neural networks trained on pairs of noisy and clean images.
# Add strong synthetic noise noisy = random_noise(original, mode='gaussian', var=0.04) noisy = random_noise(noisy, mode='s&p', amount=0.05) # extra salt & pepper
Parameters: - image: numpy array (grayscale or color) normalized to [0,1] or [0,255] - sigma: estimated noise standard deviation (used for wavelet threshold) - h: non-local means filter strength (larger = stronger denoising) - wavelet: wavelet type for thresholding
# 4. Median filter (removes any remaining salt-and-pepper noise) denoised = cv2.medianBlur((denoised * 255).astype(np.uint8), 3).astype(np.float32) / 255.0
Using "Max Denoise" in AI upscaling carries the highest risk of generating artifacts:
This is as close to “max denoise” as you can get without manually tuning per image.