//=time() ?>
Step 9: Trough of Disillusionment. Code does what it's supposed to, but there are enough parameters that most guesses look like crap... Here's garbage1003.png and garbage2047.png attached.
EVERY 👏 SINGLE 👏 TIME 👏 WITH 👏 NEURAL 👏 SYNTHESIS
Photorealistic *cough* Facial Expression Synthesis by the Conditional Difference Adversarial Autoencoder https://t.co/DE4q41pqWY #AI #ML
Converting sketches to 3D face and caricature models: https://t.co/faHYQRRb9T It's using deep learning again, what else these days? #AI #ML
Here's the breakdown of the first image during coarse-to-fine patch-matching. It's pretty cool to see actually: #DeepLearning
I haven't implemented the most computationally intensive part yet, maybe it'll save the day. Results are OK, fragility to scale is worst...
Getting more comfortable with the #NeuralAnalogy algorithm, but I can safely say that parameter hell is still there—maybe worse than before.
A critical part of working on computer vision or neural synthesis algorithms is designing incremental test cases. For example this pair:
Procedurally Generated Medieval Cities https://t.co/togGPiwB9k Love the results & style! Starts with voronoi mesh then more artisanal steps.
Transforming photos into mosaics with geometric shapes. https://t.co/fglvZzxhog
Secret to make #DeepDream not look kitsch? Use lower-level convolution layers: https://t.co/hfk8yHUC5J