Oliver presented a very nicely digestable and focused slide deck on his works on making the carbon impact of images in the web smaller.
-
Idea: Images do not need all information transmitted. They can be lossless compressed , sent, and then upscaled on the client.
-
Comparison to the human brain which can even derive the original pictures shape / meaning from very pixelated images.
-
Transposed Convolutional Neural Network can do this job perfectly.
-
After some technical deep dive (read the full slides
) and some assumptions of the cost of energy for network transmissions he presents the actual savings in terms of the inference / download comparison: 25% - 63%
If you liked the Sneak Peak you can find the full slides on the ecoCompute website: https://www.eco-compute.io/files/slides_2025/01_Thursday/02_DIM/09_Winks_Generative_Networks.pdf