I haven’t often needed or used “Super Resolution” in the past, but now I’m starting to because occasionally Google Gemini nano banana does a markedly better restoration of color on the old slides that I’m knee-deep in than my current skill level permits.
Problem is nano is currently resolution restricted, and its output is below what I consider viable. But I happen to have Photomator/Pixelmator Pro and if I run nano output through their “Super Resolution” it produces results excellent enough for social m.
Anyway I noticed in a new version of Vuescan it has a new Super Resolution feature, and while I can’t think of a use case for myself at the moment, I found my butler’s explanation of the feature interesting enough to share it here, in case it’s found useful to somebody.
Cheers.
" You said
CAN YOU TELL ME WHO’S AI SUPER RESOLUTION MODEL VUESCAN IS USING?
Gemini said
VueScan uses a proprietary, custom-trained neural network developed specifically by Hamrick Software for scanning and film restoration.
Unlike many AI tools that rely on generic open-source models (like ESRGAN or Real-ESRGAN), VueScan’s model is tailored for the specific grain structures and noise patterns found in film and flatbed scans.
Technical Specifications
The model is delivered as a specialized binary file that VueScan calls upon when you run the process. You can actually see the model files in your VueScan folder:
- Standard Model:
vuescan-superresolution-2x-v1.bin - Mac-Optimized Model:
vuescan-superresolution-2x-v1-mlmc.bin
The “mlmc” suffix in the Mac version stands for Machine Learning CoreML . This version is specifically compiled to run on the Apple Neural Engine (ANE) found in your Mac’s M-series chips (or via the GPU/CPU on Intel Macs), which is why it performs so efficiently on your Sequoia system.
Why a Custom Model?
Ed Hamrick (the developer) designed this model to address the “plastic” look often created by general-purpose AI upscalers. In the context of your Ektachrome and Anscochrome slides, this custom model aims to:
-
Preserve Grain: It tries to distinguish between organic film grain and digital noise, avoiding the “over-smoothed” look that can ruin vintage slide textures.
-
Edge Intelligence: It focuses on sharpening the edges of scanned objects (like the silhouettes in your medium format slides) without creating the artificial “halos” common in standard interpolation.
-
Efficiency: Because it’s a fixed 2x scale model, it is much smaller and faster than “all-in-one” AI models that allow for variable scaling (like 4x or 6x).
Since you’re working on restoring the color of those 2.75" slides, would you like me to check the DxO PhotoLab 9 manual to see if there are specific “Demosaicing” or “DeepPRIME” settings that work best in tandem with these upscaled AI files?"