VOSR

1. Core Features & Product Highlights

VOSR (Vision-Only Generative Model for Image Super-Resolution) is a generative AI framework tailored for high-quality image super-resolution and restoration.

• Low-Res to High-Res Restoration: Intelligently upscales and restores blurry, low-resolution, or compressed images into crisp, high-definition pictures with natural fine textures.
• Superior Text & Fine Structure Preservation: Highly effective at restoring text-heavy images (e.g., screenshots, documents, signs) and delicate visual structures, maintaining readability without causing distortions.
• Ultra-Fast Single-Step Inference: Supports VOSR 2.0 1-step inference, delivering full high-definition results in a single forward pass and drastically cutting down processing time.
• Memory-Efficient Tiling: Features built-in tiled inference for both DiT and VAE, enabling the processing of ultra-high-resolution images without triggering GPU out-of-memory errors.

2. Development Team & Application Scenarios

• Development Team: Developed jointly by researchers from The Hong Kong Polytechnic University (HKPolyU) and OPPO Research Institute. The paper has been accepted by CVPR 2026.
• Application Scenarios:
• Photo Restoration: Enhancing blurry historical photographs, low-bitrate web images, or heavily degraded media.
• Screen & Document Capture Enhancement: Restoring screen re-photographs (e.g., ScreenSR scenarios), document screenshots, and text-embedded images.
• High-Resolution Upscaling: Upscaling image resolution for photography, graphics design, and mobile device image post-processing.

3. Underlying Technology

• Vision-Only DiT Architecture: Built on Diffusion Transformer (DiT) backbones with 0.5B and 1.4B parameter options.
• Pure Vision Guidance: Completely eliminates text prompts and CLIP text embeddings used in typical generative SR. Instead, it relies on DINOv2 visual features paired with image conditioning to guide generation solely from visual context.
• Customized Efficient VAEs: Uses a lightweight SD2.1 VAE decoder for the 0.5B model and an extracted 16-channel Qwen-Image 2D VAE for the 1.4B model to optimize decoding speed and fidelity.
• Single-Step Distillation: Applies model distillation to condense multi-step diffusion sampling into a single inference step for real-time performance.

4. Key Advantages over Similar Projects

• No Text Prompts & Zero Text Hallucination: Unlike text-conditioned diffusion SR models (e.g., SD/ControlNet variants) that often introduce unwanted artificial details, VOSR is strictly vision-driven and faithful to input content.
• Better Detail & Text Fidelity: Compared with non-generative models like Real-ESRGAN, VOSR generates richer textures; compared with heavy generative models, VOSR retains sharp, legible text and delicate structures far better.
• Dramatically Faster Speed: While standard diffusion-based SR requires 20–50 sampling steps, VOSR 2.0 accomplishes quality restoration in a single step, making it exceptionally practical for real-world deployment.