Model Training ‐ Comparison ‐ [Clip Skip]

_{Models | Logs | Graphs | Configs}

---

Clip Skip (CS) determines which layer from the end of the CLIP model vectors will be sent to the U-Net. There are a total of 12 layers, and by default, the vectors are sent from the last layer. However, in leaked Novel AI checkpoint, the vectors are sent from the second-to-last layer. Due to the merging of this checkpoint with many others, CS = 2 went around the world.

That's what the guides, documentation and Google say. However, even if we try to change CS at the base SD1.5 checkpoint, it will still influence the result, although it would seem that it should not.

Compared values:

• 1 - D,

• 2 - B.

---

_DLR(step)

_Loss(epoch)

Neither DLR nor loss are affected.

---

The grids will be a little unusual this time: we will set CS not only when training the model, but also when generating images.

---

---

CONCLUSION

It seems that the difference is not significant, but I prefer to use 2 both during training and generation.

---

Next - Model Training ‐ Comparison - [Precision]