This paper presents a comprehensive methodology for designing practical and highly undetectable steganographic systems for real digital media. The core principle is to minimize distortion by using efficient coding algorithms, where distortion is defined as a weighted difference of extended feature vectors used in steganalysis. This approach allows the steganographic model to be "preserved" by steganalysts, making the system undetectable even for large payloads. The framework can efficiently handle high-dimensional models, even when the dimensionality exceeds $10^7$. The paper introduces HUGO, a new embedding algorithm for spatial-domain digital images, which outperforms LSB matching in terms of message length and security. On the BOWS2 image database, HUGO allows the embedder to hide 7 times longer messages with the same level of security. The method is based on the principle of minimal impact embedding, which separates the image model from the coding algorithm, allowing for the simulation of optimal coding and the comparison of image models without coding effects. The paper also discusses the use of SPAM features to derive embedding costs and the identification of detectable parts of the models. The security of HUGO is verified through experiments, showing that it can hide messages with a relative payload of 0.3 bpp at a detection error of 40%, compared to 0.04 bpp for LSB matching.This paper presents a comprehensive methodology for designing practical and highly undetectable steganographic systems for real digital media. The core principle is to minimize distortion by using efficient coding algorithms, where distortion is defined as a weighted difference of extended feature vectors used in steganalysis. This approach allows the steganographic model to be "preserved" by steganalysts, making the system undetectable even for large payloads. The framework can efficiently handle high-dimensional models, even when the dimensionality exceeds $10^7$. The paper introduces HUGO, a new embedding algorithm for spatial-domain digital images, which outperforms LSB matching in terms of message length and security. On the BOWS2 image database, HUGO allows the embedder to hide 7 times longer messages with the same level of security. The method is based on the principle of minimal impact embedding, which separates the image model from the coding algorithm, allowing for the simulation of optimal coding and the comparison of image models without coding effects. The paper also discusses the use of SPAM features to derive embedding costs and the identification of detectable parts of the models. The security of HUGO is verified through experiments, showing that it can hide messages with a relative payload of 0.3 bpp at a detection error of 40%, compared to 0.04 bpp for LSB matching.