Skip the diffusion queue: super-resolution in one to five steps

What it does InvSR is a diffusion-based image super-resolution system built around Stable Diffusion Turbo. It trains a deep noise predictor to estimate an intermediate diffusion state from a low-resolution input, then samples forward to a high-resolution output in anywhere from one to five steps. The authors report that even a single step produces results comparable to recent state-of-the-art methods.

The interesting bit The core idea is Partial Noise Prediction: instead of starting from pure Gaussian noise and iterating dozens of times, the model learns to guess the optimal noise map that fast-forwards the diffusion trajectory to a useful midpoint. That turns the generative prior into an adjustable upscaler with a quality-speed dial ranging from one to five steps.

Key highlights

• Arbitrary sampling steps from 1 to 5; the paper claims single-step quality rivals recent SOTA.
• Targets both real-world photographs and AI-generated content (SDXL, Flux).
• Built on Stable Diffusion Turbo and the Hugging Face diffusers library.
• Accepted to CVPR 2025.
• Ships with Gradio, Replicate, Hugging Face, and Colab demos.

Caveats

• Reproducing quantitative results on some datasets requires a --color_fix wavelet flag, so out-of-the-box colors may drift without it.
• Large upscales (e.g., 1K→4K) require tiled inference with a chopping_size parameter to fit in memory.
• Released under the NTU S-Lab License 1.0, which carries redistribution restrictions stricter than standard open-source licenses.

Verdict A strong candidate if you want diffusion-quality upscaling with an explicit speed/quality trade-off and already run a PyTorch stack. Less appealing if you need a permissive open-source license or a memory-agnostic drop-in tool.