# Overfitted Codecs: What if the decoder was the image? Talk given at the [Kitchener-Waterloo Video Technology Meetup](https://www.youtube.com/channel/UCU5ZJDoB9T3oR-HuJDAS6YQ) in April 2026. <div style="display: flex; justify-content: center;"> <iframe width="560" height="315" src="https://www.youtube.com/embed/lkpyY7ClRkc?si=DfWA_GLDd9glBqEZ" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe> </div> ## Abstract What if the best way to compress an image was to build a tiny neural network that only works for that one image, and throw away the pixels entirely? This talk traces a journey from the basics of neural networks to the state-of-the-art in learned image compression; through Implicit Neural Representations, SIRENs, COIN, and CoolChic. Along the way it challenges one of machine learning's cardinal rules: that overfitting is always the enemy. ## Scope and intent Learned compression is often presented as a black box; neural networks that somehow outperform traditional codecs on benchmarks, with little explanation of why. This talk focuses on: - what it actually means to represent an image as a function - why overfitting, normally a failure mode, becomes a design goal - how gradient descent replaces the encoder - where overfitted codecs sit relative to the broader trajectory of video compression The emphasis is on building intuition, not surveying benchmarks. ## Slides 📄 [Download slides (PDF)](https://files.trud.ca/OverfittedCodecs.pdf) ## Key Takeaways - **A bit and a Shannon are not the same thing.** A bit measures storage. A Shannon measures surprise. The goal of compression is to make the bits match the Shannons. - **An image is a function, not a grid.** f(x, y) → color. Storing the function instead of the samples is the insight that unlocks everything else. - **Overfitting is a feature, not a bug.** When you only need to compress one image, perfectly memorizing it is exactly the right strategy. - **Gradient descent is the encoder.** In overfitted codecs there is no separate encoding step, optimization IS compression. The loss function defines quality. - **The bitstream is a checkpoint, not the end.** Unlike traditional codecs, you can always resume encoding, refine the result, and push quality further on demand. ## References - Sitzmann et al., [Implicit Neural Representations with Periodic Activation Functions (SIREN)](https://arxiv.org/abs/2006.09661), NeurIPS 2020 - Dupont et al., [COIN: COmpression with Implicit Neural representations](https://arxiv.org/abs/2103.10943), ICLR Workshop 2021 - Ladune et al., [COOL-CHIC: Coordinate-based Low Complexity Hierarchical Image Codec](https://arxiv.org/abs/2212.05458), ICCV 2023 ## Feedback and discussion Questions, disagreements, and follow-up thoughts are welcome, especially from people working on traditional pipelines who found something to push back on. If you're interested in a more technical version of this material — covering rate-distortion theory, perceptual quality metrics, or where learned compression realistically fits in a production pipeline — reach out via [LinkedIn](https://www.linkedin.com/in/luctrudeau) or through the [Consulting](https://trud.ca/Consulting) page.