You Only Landmark Once: Lightweight U-Net Face Super Resolution with YOLO-World Landmark Heatmaps

2026-05-15 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

Researchers propose a lightweight U-Net based architecture for face image super-resolution, designed to reconstruct 128x128 facial images from severely degraded 16x16 inputs, achieving an 8x magnification. The method introduces a novel auxiliary-training-free supervision strategy that utilizes heatmaps generated by YOLO-World, an open-vocabulary object detector, to localize key facial features like eyes, nose, and mouth. These heatmaps are converted into spatial weights, forming a heatmap-guided loss that emphasizes reconstruction errors in semantically important regions. Unlike prior methods requiring dedicated landmark or alignment networks, this approach reuses detector outputs as supervision, maintaining an efficient training and inference pipeline. Experiments on the aligned CelebA dataset demonstrate that the proposed loss consistently improves quantitative metrics and produces sharper, more realistic reconstructions without adversarial training or increased computational cost.

Key takeaway

For research scientists developing efficient face super-resolution models, this work demonstrates that integrating detection-driven priors from general-purpose models like YOLO-World can significantly improve reconstruction quality without adding computational overhead or requiring dedicated landmark networks. You should consider adapting this heatmap-guided loss strategy to enhance detail in critical regions, especially when targeting resource-constrained or real-time applications.

Key insights

YOLO-World heatmaps can guide lightweight U-Net face super-resolution without auxiliary training.

Principles

• Leverage existing detectors for supervision.
• Prioritize reconstruction errors in key regions.
• Lightweight networks can achieve extreme upscaling.

Method

Generate heatmaps from YOLO-World detections of facial features, apply edge detection and spatial fading, assign class-specific weights, and integrate into a weighted pixel loss to guide a U-Net architecture for 8x face super-resolution.

In practice

• Use YOLO-World for zero-shot landmark detection.
• Apply Gaussian/Inverse Gaussian fades for feature emphasis.
• Combine pixel, perceptual, heatmap, and LPIPS losses.

Topics

• Face Super-Resolution
• Lightweight U-Net Architecture
• YOLO-World
• Landmark Heatmaps
• Heatmap-Guided Loss

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Surflo: Consistent 3D Surface Flow Model with Global State — Surflo uses a global latent state and flow matching to reconstruct arbitrary-resolution 3D surfaces from variable image inputs.
• Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models — YOLO26 unifies real-time vision tasks with NMS-free inference and advanced training, improving accuracy and efficiency.
• A Scale-Adaptive Framework for Joint Spatiotemporal Super-Resolution with Diffusion Models — A scale-adaptive framework reuses one architecture for spatiotemporal super-resolution across diverse scaling factors.
• DVANet: Degradation-aware Visual-prior Alignment Network for Image Restoration — DVANet unifies image restoration by combining degradation-aware consistency with DINOv3-guided visual priors via deep unfolding.
• Learning Where to Embed: Noise-Aware Positional Embedding for Query Retrieval in Small-Object Detection — Selectively embedding positional information in foreground regions improves small-object detection and reduces model complexity.
• Honey, I Shrunk the Arc de Triomphe! — A new dataset and method mitigate scale-collapse in monocular depth estimation for distant, unconstrained scenes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.