SCEESR: Semantic-Control Edge Enhancement for Diffusion-Based Super-Resolution

TL;DR

SCEESR addresses Real-ISR challenges by enhancing a one-step diffusion SR model with semantic edge guidance via a ControlNet and a gated fusion of Canny and HED edges. A LoRA-tuned diffusion backbone preserves efficiency, while a hybrid loss including an adaptive edge term enforces both pixel accuracy and geometric fidelity. Quantitative and qualitative results show improved structural integrity and perceptual realism with only modest increases in computation. This approach offers a practical path to high-quality SR in real-world scenarios where speed and detail preservation are both critical.

Abstract

Real-world image super-resolution (Real-ISR) must handle complex degradations and inherent reconstruction ambiguities. While generative models have improved perceptual quality, a key trade-off remains with computational cost. One-step diffusion models offer speed but often produce structural inaccuracies due to distillation artifacts. To address this, we propose a novel SR framework that enhances a one-step diffusion model using a ControlNet mechanism for semantic edge guidance. This integrates edge information to provide dynamic structural control during single-pass inference. We also introduce a hybrid loss combining L2, LPIPS, and an edge-aware AME loss to optimize for pixel accuracy, perceptual quality, and geometric precision. Experiments show our method effectively improves structural integrity and realism while maintaining the efficiency of one-step generation, achieving a superior balance between output quality and inference speed. The results of test datasets will be published at https://drive.google.com/drive/folders/1amddXQ5orIyjbxHgGpzqFHZ6KTolinJF?usp=drive_link and the related code will be published at https://github.com/ARBEZ-ZEBRA/SCEESR.

Figures (5)

• Figure 1: The training process of SCEESR. The LQ image is passed through a frozen encoder $E_\theta$, a LoRA finetuned diffusion network $\epsilon_\theta$ and a frozen decoder $D_\theta$. Text prompts, generated by a prompt encoder, are fed into both the diffusion network $\epsilon_\theta$ and a semantic control gate. Simultaneously, Canny edge maps and HED edge maps are extracted from the LQ image using a Canny detector and an HED detector, respectively. The outputs from the two ControlNets—corresponding to the edge maps—are fused in the semantic control gate under the guidance of the text prompts. The fused control information is then injected into each layer of the diffusion network $\epsilon_\theta$. During inference, only the diffusion network $\epsilon_\theta$, the two ControlNets, and the parameters of the semantic control gate are utilized.
• Figure 2: The figure presents a comparative visualization of edge detection results from the same input image (left). The middle image shows the output of the Canny detectorcanny2009computational The right image displays results from the HED detectorxie2015holistically
• Figure 3: Vision comparisons of different GAN-based and DM-based SR methods on Lincoln's face.
• Figure 4: Vision comparisons of different GAN-based and DM-based SR methods on a frog's toes.
• Figure 5: Qualitative comparison with state-of-the-art methods on both synthetic and read-world benchmarks between the results of the first phase of SCEESR with only L2 loss and LPIPS loss(the first row) and the results of the second phase of SCEESR with semantic control edge enhancement ControlNet and AME loss(the second row).