The following sections illustrate the development of the project over time. The following models are considered legacy and while certain imagery is of a high quality these are not classified as the best works, however, we feel it is important to show how the project started and developed over time and images will be purchasable from these early models for posterity in time.
We added nVidia V100 cards to our solution and again utilised a source set of paintings from abst.ract.me. The network generated images of size 256×256 with modified network parameters.
This network was classified as the first truly successful network since it started generating images that illustrated brush marks, canvas texture and shows true character.
This network failed to produce works beyond 128×128, the nVidia M10 cards were unable to generate 256×256 images due to memory limitations. The network generated images were again from a source set of paintings from abst.ract.me, whilst interesting images were generated they were still at 128×128.
This network produced good results at a larger resolution of 128×128 using the same algorithm as Daniella. With the new hardware, we were able to generate the 128×128 images of the Daniella network.
This network produced poor results at low resolution. We took the Cain model and data, updated the settings and ran the model on the new hardware which has significantly more resources, however, the results were still poor. We dubbed these results the Neon series.
This network produced good results at low resolution. We modified the learning rates and the number of source images but kept input sizes the same as Alberto. Coreix moved our deployment of Open Nebula and built us a private cloud solution using Open Stack with nVidia Tesla M10 cards so that we could run multiple tests. Daniella produced interesting results, if at low resolution.
This network produced poor results. We kept the Convolutional Neural Network from Alberto and applied it to a different source which was the photographs of SGD Photos. The network failed to produce good results from the photographs. We felt at this point we needed more powerful hardware and a new environment.
The Becky network produced some poor results at low resolution. The network generated images was again from a source set of paintings from abst.ract.me. The images generated were from the same sources but learning rates were changed and batch sizes modified. The experimentation produced relatively poor results.
The first network to produce some very basic results at low resolution. Alberto generated images from a source set of paintings from abst.ract.me. The images generated were from a relatively small set and generated 64 x 64 output images. At this point, we had progressed from the test environment to a small Open Nebula deployment and utilised the nVidia RTX 3060 series cards.
Our first network was developed. These were generally low-resolution images which are more developed as a proof of concept. We were utilising a test environment with dual nVidia RTX 1080 series cards for image generation.