📸

Image Upscaling - SMFANet+

🛠

Fast, Lightweight Image Upscaling

A small, efficient image upscaler that takes a low-resolution photo and hands back a 2×, 3×, or 4× larger version in as little as 0.1 – 3 seconds. SMFANet+ is the boosted variant of SMFANet — a lightweight super-resolution model from Nanjing University of Science and Technology (ECCV 2024) that placed 2nd and 3rd in the Parameters and FLOPs sub-tracks of NTIRE 2024 ESR, meaning it produces excellent upscaled quality at a very small computational cost.

What it does

SMFANet+ makes images bigger and sharper. Feed it a low-resolution photo, pick a scale factor (2×, 3×, or 4×), and it produces an enlarged version with fine detail restored. It is a lightweight super-resolution model — meaning the model itself is small and fast, so upscaling takes only a fraction of a second on typical hardware while still producing quality that competes with much heavier models. The "+" variant is a stronger sibling of the base SMFANet architecture, tuned for higher output quality while staying in the lightweight class. Ideal for batch upscaling, real-time or near-real-time workflows, and deployments where compute or memory is constrained.

Problem it solves

  • Fast batch upscaling – Small model = fast inference, so you can upscale many images without a queue backing up
  • Low-latency workflows – Sub-second to few-second turnaround makes this practical for interactive or near-real-time pipelines
  • Runs on modest hardware – Very small parameter count and low FLOPs, so it works well on constrained GPUs (or even without one)
  • Choice of 2×/3×/4× – Pick the exact enlargement factor for the job, from mild upscaling to aggressive 4×
  • Small / low-res source photos – Thumbnails, chat images, archived scans, old web images, screenshots
  • Video frame upscaling – Fast enough to be practical for upscaling frames pulled from video
  • TensorRT-ready – The upstream project ships a TensorRT export path for even faster deployment

Input/Output

  • Input: Any RGB photo (small, blurry, or low-resolution)
    • image
  • Parameters:
    • Upscale ratio (dropdown, default 2) — pick 2, 3, or 4 to control how much larger the output should be
      • image
  • Output: Sharp, clear, high-resolution version (bigger and clearer) (2x, 4x bigger)
    • image

Accuracy & Speed

Speed
~0.1 – 3 seconds per image, depending on input size and scale factor
Efficiency
Placed 2nd and 3rd in the Parameters and FLOPs sub-tracks of the NTIRE 2024 Efficient SR (ESR) Challenge — one of the most parameter- and compute-efficient SR models in its quality class
Model size
Lightweight — designed for constrained deployments
Scale factors
2×, 3×, 4× (single model per scale)
Deployment
Standard PyTorch inference + optional TensorRT export for further speed-ups
Recognition
Published at ECCV 2024

How SMFANet+ compares

SMFANet+ sits in CNAPS Studio's Image Upscaling lineup as the fast, lightweight option — the one to reach for when latency, throughput, or compute budget matters more than squeezing out the absolute maximum PSNR. Its siblings in the studio target different points on the frontier:

  • Swin2SR / SwinIR — heavier transformer models, higher peak quality on classical benchmarks, slower to run
  • PiSA-SR — real-world / degradation-focused upscaler for messy input
  • SMFANet+ (this node) — smallest and fastest of the group; best when you're doing batch or interactive work and don't need the last dB of PSNR

All run inside CNAPS Studio — compare them side-by-side.

‣
Technical Details (for engineers)

Model Source

Compliance & Provenance

Provider
Open-source — IMAG Lab, Nanjing University of Science and Technology
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Minimal Risk
Art. 50 transparency
Not applicable
Region availability
Available globally
Training data summary
DIV2K (per the upstream training configs); see provider repo for details

For more on how we classify models and mark outputs, see our AI Policy.

Run it now