Model Training ‐ Comparison ‐ Final

_{Models | Logs | Graphs | Configs}

---

Ok, now we'll make final comparison:

• Base Settings (aka Base Settings 1),

• Base Settings + Growth Rate = ∞ (aka Base Settings 2, aka model from growth rate comparison),

• Base Settings 2 + Regularisation with Unsplash Photos at Max Resolution using RV2.0 model (aka Intermediate Settings 1, aka model from regularisation comparison),

• Base Settings 2 + Regularisation with Unsplash Photos at Max Resolution using 28D28 model (aka Intermediate Settings 2, aka model from regularisation comparison),

• Final Settings + Regularisation with Unsplash Photos at Max Resolution using RV2.0 model (aka Final Settings 1),

• Final Settings + Regularisation with Unsplash Photos at Max Resolution using 28D28 model (aka Final Settings 2).

---

Billie Eilish

_Loss(epoch)

_DLR(step)

---

---

Karina Istomina

_Loss(epoch)

_DLR(step)

---

---

And I chose the most interesting results on both characters.

---

So, for both characters we achieved pretty good results with both intermediate and final settings. It still quite depends on a checkpoint you're using for generation, but anyway. It means that the major and crucial thing that drasticly improves similarity and quality is regularisation. Everything else is minor. You can change almost everything within reasonable limits and smart optimizer will try his best to not allow you to spoil the results. But as you could see it's still possible to achieve good results even without regularisation though it requires a lot more tries.

If you don't feel like manually filling in the settings following the guide, you can use my configuration files. Here's one and here's another. They only differ in the checkpoint used for training. Don't forget to replace the placeholders with your own values! If something doesn't work, go through the list step by step and change the parameters where multiple possible values are indicated.

---

Next - Image Generation