<p>This work proposes a lightweight and dataset-independent framework that formulates facial texture and emotion transfer as a multi-scale optimization problem rather than a generative learning task. Given an input and a reference facial image, the method aligns facial regions using landmark-based warping and extracts high-frequency texture components through sparse representations with a fixed DCT dictionary, followed by cartoon-texture separation. The extracted texture and emotional cues are selectively fused using a gradient descent optimization scheme guided by a combined loss function incorporating reconstruction fidelity, total variation regularization, and color distribution consistency, enabling operation on facial images of arbitrary resolution. Experimental evaluations on the TUFTs RGB Emotion and RaFD datasets demonstrate that the proposed method consistently increases the probability of the target emotional state by an average of 35%, leading to an emotional change in 91% and 87% of the simulations on the RaFD and TUFTs datasets, respectively, as measured by an independent emotion classifier. The method primarily enhances the presence of target emotional cues while preserving facial identity and minimizing visual artifacts, rather than strictly enforcing categorical emotion replacement. Extensive qualitative and quantitative comparisons with state-of-the-art approaches further indicate that the proposed method produces stable, photo-realistic results with reduced distortion, offering a robust alternative for facial texture and emotion transfer under limited data and computational constraints.</p>