Visualizations were created for the following models in decreasing order of loss:
- Basic+masking-dynamic (loss: 0.00691)
- Basic+masking-uniform (loss: 0.00671)
- Basic (loss: 0.00671)
- Basic+masking-4 (loss: 0.00651)
- Basic+masking-16 (loss: 0.00646)
- Basic+masking-32 (loss: 0.00639)
- Basic+masking-8 (loss: 0.00638)
As a reminder, as stated here, the baseline loss is set at 0.01100. All models have successfully surpassed the baseline and have been able to restore a portion of the original image's color while simultaneously upscaling it from 64x64 to 178x178 pixels.
Before proceeding, it is recommended that you read this note on visualizations for better understanding.
It is interesting to note that models trained on larger masked areas, such as a 4x4 or 8x8 masking grid, tend to better preserve textures. This could potentially be utilized in the future to enhance the results.
(To understand the concept of local region restoration, please refer to the following note.)
The results are not particularly impressive, but the ability to reconstruct a local region is an interesting feature. With more accurate models, the results would be much better.