Controllable Image Enhancement Using Adaptive Scaling in Deep Super-Resolution Networks

Most image enhancement models operate at fixed scaling factors typically $2 \times, 4 \times$, or $8 \times$ without allowing users to control over how much detail they want to recover. As a result, outputs are often either over-smoothed or insufficiently enhanced. Here, we present a deep learningbased method that introduces a continuous and usercontrollable enhancement mechanism, allowing users to specify the desired restoration level on a normalized scale from 0.00 to 1.00. A value of 0 applies no enhancement, while 1 trigger full super-resolution. This adaptive design eliminates the need for multiple models for different upscale scenarios or retraining at different scales. This adaptive design is lightweight, resourceefficient, and well-suited for real-time applications such as web applications, mobile photography, live video enhancement, and device-adaptive media. The network dynamically adjusts its enhancement strength based on the scaling input, offering a smooth trade-off between visual quality and computational cost. Extensive experiments demonstrate that the proposed method delivers high perceptual quality and consistent performance across varying enhancement levels, outperforming traditional fixed-scale approaches in both efficiency and flexibility.