ASASR turns blurry photos sharp without AI hallucinations

ASASR is an open-source image super-resolution tool built on top of FLUX.1-dev. It upscales blurry, low-quality images by 4x while keeping the model from inventing weird, hallucinated artifacts that often plague generative upscalers.

Traditional diffusion-based upscalers often make photos look sharp at the cost of fidelity, fabricating fake skin pores or mangling text. ASASR is designed to fix that exact problem, delivering crisp reconstructions that preserve legible letters and clean object boundaries.

How it works

The project uses a method called Adversarial Sobolev Alignment to keep high-frequency structures grounded in reality.

Basically, it replaces standard Gaussian noise with a spectrally shaped noise kernel that matches natural image distributions. On top of that, an adversary network creates realistic artifact-heavy negative samples on the fly, which the model uses to steer away from bad reconstructions through frequency-weighted Direct Preference Optimization.

During inference, it runs a dual-LoRA pipeline. A base SR LoRA weighing about 885 MB provides the reconstruction prior, while a smaller 111 MB DPO LoRA cleans up artifacts and refines structural fidelity on top of the FLUX backbone.

In standard benchmarks, ASASR ranks at the top of the DIV2K-Val leaderboard for Real-ESRGAN 4x degradation and sits at number three on RealSR.

Specs and how to try it

If you want to run it locally, the main hurdle is FLUX itself.

You need the 24 GB FLUX.1-dev base model alongside the two LoRA checkpoints, which download to about 1 GB total. The code runs on Python 3.10, PyTorch 2.x, and diffusers 0.30 or newer with BF16 precision. A single A100 GPU handles the bundled inference examples in about a minute.

If you just want to see how it looks on your own photos, you do not need a high-end rig. Hugging Face hosts an official web demo on Spaces where you can upload an image and run the 4x pipeline directly in your browser.

Keep in mind that the current public checkpoints are specifically trained for 4x scaling from 128 to 512 pixels. The repository is released under a CC-BY-NC-4.0 license, meaning it is strictly for non-commercial research right now.

Why it matters

If you frequently restore vintage photos or need to clean up blurry assets where text and fine lines matter, ASASR is a serious upgrade over older generative tools like SUPIR. Those models produce jaw-dropping sharpness, but they often hallucinate completely made-up details where they should not.

That said, using a massive model like FLUX for image upscaling is computationally heavy. ASASR is fantastic for high-stakes offline restoration, but its hardware appetite means it is not ready to replace lightweight, real-time upscaling pipelines just yet.