The work described in this article is the result of a collaborative effort across multiple teams at Multiverse Computing. You can find the full list of contributors at the end of this article.
Modern computer vision models face a recurring challenge: as they grow more sophisticated and capable of solving problems, their operational footprint often makes them hard to deploy in cost-sensitive or edge environments.
In this case, the project started with a proprietary computer vision model and a clear deployment target: edge hardware with tight memory and energy constraints. Lots of times, running models on constrained edge hardware is simply not viable.
At Multiverse Computing, we set out to change that by asking a single question: Can we compress a high-fidelity model by 95% and preserve accuracy while improving speed and energy consumption? The stakes were concrete: without compression, deployment was off the table entirely.
Using our quantum-inspired compression techniques, we moved beyond theoretical benchmarks to achieve a result that redefines deployment constraints.
The Engineering Snapshot: Efficiency Without Sacrifice
Before diving into the methodology, here is the high-level impact of the optimization process, achieved without a tangible effect on accuracy metrics:
The Strategy: Quantum-Inspired Tensorization
To achieve these results without redesigning the architecture from scratch, we applied a dual-technique approach:
- Structural Tensorization: Imagine a city cataloged brick-by-brick (millions of entries) versus noting that the city follows three specific blueprints (stored efficiently). Instead of treating layers as dense parameter blocks, we restructured them into compact tensor representations. This allows the model to express complex transformations with significantly fewer parameters.
- Model Quantization: As a complementary step, we optimized how network weights are represented, further accelerating runtime performance.
Higher throughput means your hardware processes nearly twice as many images in the same time window. And not only we are processing twice as many images, but each of them needs almost half the energy to be processed (measured as the average joules consumed per forward pass under controlled conditions).
At scale, this translates directly to reducing cost per inference and thermal constraints on edge devices, enabling novel deployments that could not be possible in the past.
The city blueprint analogy for tensorization.
A neural network's parameters are like a city catalogued brick by brick: millions of individual entries, stored without regard for repetition. Tensorization identifies the underlying blueprints already implicit in that structure: a small set of compact representations that reconstruct the same information far more efficiently. The structure was always there; tensorization simply finds and stores it that way.
This process was iterative and progressive. By optimizing layers in stages and validating quality at each step, we avoided the "risky jump" that often leads to accuracy collapse. Once we found the three blueprints that lied beneath the millions of bricks, the model went from 258 MB to under 9 MB, and ran nearly twice as fast without compromising the output quality.
The "Hidden" Benefit: Better Generalization
Compression acted as a regularizer by reducing the model's capacity to memorize all the bricks one by one, and it forced out an overfitting that the standard training had masked.
We performed a retraining with augmented data to assess this, and it produced a model that was not only smaller and faster, but also more robust on out-of-distribution inputs.
Data Integrity: Does it still work?
To validate quality at 95% compression, we needed metrics that capture both pixel-level fidelity and perceptual structure. We evaluated against two industry standards across 20,780 samples:
- PSNR (Peak Signal-to-Noise Ratio) measures how much information is lost between the original and the reconstructed image. The higher the value, the closer the output is to the original. It is expressed in decibels (dB), and in practice, values above 40 dB indicate reconstruction quality where differences are invisible to the human eye.
In our evaluation, PSNR dropped from 42.93 dB to 40.84 dB. While this is a 2.09 dB change, it remains squarely within that high-quality range, meaning the compressed model produces outputs that are, for all practical purposes, indistinguishable from the original. - SSIM (Structural Similarity Index) goes beyond pixel-level differences and evaluates whether the reconstructed image preserves the perceptual structure of the original: its contrast, luminance, and texture patterns. Scores range from 0 to 1, where 1 means perfect structural agreement. In our evaluation, SSIM barely moved, shifting from 0.98 to 0.97. Effectively identical, and well above any threshold that would indicate meaningful degradation.
Redefining the Limit
The redundancy in large neural networks is far greater than intuition suggests. By reducing the cost of running intelligence, we enable AI to move from the lab into real-world, where hardware-constrained environments make efficiency a hard requirement.
The takeaway is clear: You don't have to trade accuracy for efficiency; you just need to find the "blueprints" already hidden within your data.
Strategic Impact: What This Means for Your Business
Reducing a model from 258.58 MB to 8.22 MB is more than a technical milestone; it is a catalyst for operational efficiency. For companies looking to scale AI solutions, this breakthrough removes traditional infrastructure barriers and accelerates time-to-value:
- Edge Mastery: Your high-tier computer vision models can now run on lightweight hardware that previously could not accommodate the original file size.
- Capacity Acceleration: By nearly doubling throughput from 42.90 to 81.93 img/s, your business gains higher system capacity on existing hardware, effectively lowering the cost per inference.
- Agile Distribution: At just 8.22 MB, models load and transfer faster, eliminating the distribution overhead that typically slows down update cycles and edge deployments.
- Critical Energy Efficiency: Cutting energy usage by ~49% (from 3.41 J/img to 1.74 J/img) enables large-scale operations in environments where thermal or power constraints are a hard limit.
These results show that model compression can materially change the economics of computer vision deployment. In this case, the compressed model became significantly smaller, faster, and more energy-efficient while preserving high-fidelity output quality across reconstruction metrics.
Acknowledgement
This work would not have been possible without the support of several people. A special thank you to Marcelo Moreno, who worked side by side with me throughout this project and whose deep understanding of tensor networks was foundational to everything we achieved here. Thank you to Alessandra Pellegata, who as my manager trusted me with this challenge and gave me the space and encouragement to pursue it. And thank you to Doğancan Gemici, Director of our team, for his continued support through it.