MossNet

MossNet is a neural network for solving super resolution problem(SR - further) preserving memory efficiency.

Abstract

The current state-of-the-art super resolution methods are based on diffusion models, which are computationally expensive. In the same time also much model utilize text-to-image models which are making inference even more complex. The MossNet aims to remove this complexity by utilizing U-Net architecture without text-to-image model.

Architecture

MossNet proposes context-aware patch-local upscaling mechanism, which allow to use single embedding "describing" image for whole upscaling process. In the same using smaller, fixed patches allow to reduce memory footprint.

Cut image into patches
Cut image into patches
Image
Image
Patches
Patches
Embedding("description")
Embedding("description")
ContextNet
ContextNet
Decode context
Decode context
For each patch
For each patch
Concat context to patches
Concat context to pa...
NxN patch
NxN patch
Initial convolution
Initial convolution
2Nx2N patch
2Nx2N patch
UNet
UNet
Add to reconstructed image
Add to reconstructed...
Text is not SVG - cannot display

Preliminary results

Experiment setup

Model training did not include noise reduction targets or deblur targets. The reason behind this choice is that the model is aimed to users of handsets and embedded devices that may want blurred images to be displayed as is and denoise algorithms are usually separate. Thus metrics are limited to LPIPS and PSNR. The model was trained of 50000 iterations on a open images datasets. The model then was distilled to XS version.

The model was evaluated on OpenImages dataset which was also used for training. Also to assess the model out-of-sample generalization ability arbitrary images from DIV2K and Unsplash were used during visual testing.

The GFLOPS parameter was measured on SR for image of 256x256 pixels.

Quantative results

ModelLPIPS↓PSNR↑GFLOPSParam Count
MossNet-XS0.13227.643.3264.2K
SRGAN⚠️0.12626.63-5.949M
StableSR@200 steps0.31123.26-approx. 200M
GuideSR0.26524.76-approx. 200M
EDSR⚠️0.13334.64-1.37M

Important notice:The above quantative results may have bias since the model was never trained on Div2K dataset, unlike models from thisthispaper. Models for which data is likely affected by this bias are marked by a warning⚠️ icon.

Though this issue is awaiting model(among others) to be specifically re-trained on DIV2K, the problem with my GPU setup would like take time to resolve. For the same reason GFLOPS is currently unavailable for most models in above table.

Visual results

DIV2K

OpenImages

Unsplash

Just random image from Unsplash.

Want to help? Or use?

Want to help mobile-first offline upscalers and enhancers come true? Want try out this model for your usecase?

FAQ

What is MossNet, in one sentence?

A memory-efficient super-resolution model that uses a U-Net architecture with context-aware patch-local upscaling, deliberately avoiding diffusion or text-to-image components to keep inference cheap enough for handsets and embedded devices.

How is MossNet different from diffusion-based super-resolution models like StableSR or GuideSR?

Diffusion-based SR models are accurate but computationally heavy, often relying on text-to-image backbones with hundreds of millions of parameters and multi-step sampling. MossNet skips both: no diffusion process, no text-to-image model. It builds a single description embedding for the whole image, then upscales fixed-size patches against that shared context — far fewer parameters and a single forward pass.

Does MossNet handle denoising or deblurring as well as upscaling?

No, by design. Training did not include noise-reduction or deblur targets, since the target users — handset and embedded-device contexts — typically want blur preserved as-is and run denoising as a separate step. Evaluation is therefore limited to LPIPS and PSNR.

How good are MossNet's results compared to other models?

MossNet-XS sits competitively on LPIPS and PSNR against much larger models (SRGAN, EDSR, StableSR, GuideSR) while using a fraction of the parameters — 64.2K versus millions. See the quantitative results table on this page for exact figures, including which comparisons carry a DIV2K-training-bias caveat.

Are the benchmark comparisons fair?

Partially caveated. MossNet was never trained on DIV2K, unlike some of the comparison models, which may bias results in either direction. Any model whose numbers are likely affected by this is flagged with a ⚠️ icon in the results table.

Can I try MossNet or use it for my own use case?

Yes — reach out by email or schedule a call using the links on this page. The project is actively looking for people interested in mobile-first, offline upscaling and enhancement use cases.

What's the license and citation for this work?

This page is licensed under CC BY-NC-ND, since the research is still in progress. The DOI is 10.5281/zenodo.15665470, linked from the Copyright section on this page.

References

Notice

As this research is still in progress, this page is left under restrictive CC license.

Copyright

This page is licensed under the CC BY-NC-ND license.

DOI: 10.5281/zenodo.15665470

Author

Rust-loving, Python-purring Rubyist with a taste for clean UI and warm naps in the sun.