GTF: Omnidirectional EPI Transformer for Light Field Super-Resolution

2026-05-06 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

GTF, an omnidirectional Epipolar Plane Image (EPI) Transformer, is introduced for light field (LF) image super-resolution, explicitly modeling horizontal, vertical, 45-degree, and 135-degree EPIs within a unified reconstruction framework. This approach addresses the underexplored diagonal epipolar geometry in existing Transformer-based LF SR methods. GTF integrates directional EPI processing, MacPI-based prior injection, adaptive directional fusion, and a topology-preserving feed-forward network. The main GTF model achieved 32.78 dB on five standard LF SR benchmarks, while a lightweight GTF-Tiny variant, designed for efficiency, reached 32.57 dB with 0.915M parameters and 19.81 GFLOPs. In the NTIRE 2026 Light Field Image Super-Resolution Challenge, GTF submissions ranked 3rd on Track 1 and Track 3, and 4th on Track 2.

Key takeaway

For research scientists developing light field super-resolution models, you should investigate incorporating omnidirectional EPI processing, including diagonal epipolar geometry, into your Transformer architectures. This approach, demonstrated by GTF, can yield substantial performance gains in fidelity and efficiency, particularly for competitive benchmarks like NTIRE challenges. Evaluate the trade-offs between model complexity and performance using variants like GTF-Tiny to meet specific resource constraints.

Key insights

Omnidirectional EPI modeling, including diagonal views, significantly enhances light field super-resolution performance.

Principles

• Explicitly model all EPI directions.
• Integrate prior knowledge via MacPI.
• Fuse directional features adaptively.

Method

GTF processes horizontal, vertical, 45-degree, and 135-degree EPIs, injects MacPI priors, adaptively fuses directional features, and uses a topology-preserving feed-forward network for LF SR.

In practice

• Consider diagonal EPIs for LF SR.
• Implement adaptive directional fusion.
• Utilize lightweight model variants for efficiency.

Topics

• Light Field Super-Resolution
• EPI Transformer
• Omnidirectional Epipolar Geometry
• NTIRE 2026 Challenge
• Transformer Architecture

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

Related on AIssential

• Gaussian Light Field Splatting: A Physical Prior-Driven Vision Transformer for Unsupervised Low-Light Image Enhancement — GLFS integrates Gaussian splatting's physical illumination modeling into a Vision Transformer for unsupervised low-light image enhancement.
• Multi-scale interaction network for stereo image super-resolution — A multi-scale interaction network enhances stereo image super-resolution by optimizing intra-view and cross-view information exploitation.
• GMBFormer: An NDVI-Guided Global Memory Bank Transformer for Urban Green-Space Extraction from Ultra-High-Resolution Imagery — GMBFormer improves urban green-space extraction by decoupling NDVI and using a global memory bank for semantic reuse.
• OmniDirector: General Multi-Shot Camera Cloning without Cross-Paired Data — The OmniDirector framework enables general multi-shot camera cloning by encoding camera motion as grid videos and training on million-scale data.
• 🍓Surface Light Tokenizer🍓 👉Apple unveils LITO a novel latent flow matching model enables... — LITO uses latent flow matching to convert images into high-quality 3D models via compact surface light field representations.
• Introducing OpenAI’s GPT-image-1.5 in Microsoft Foundry — GPT Image 1.5 offers enhanced visual fidelity, faster generation, and precise editing for enterprise visual AI workflows.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.