muooon
/

EmoNAVI

@@ -1,40 +0,0 @@
-TensorBoard
-Gray: emonavi, Light blue: AdamW
-As the graph shows, emonavi's progression is slightly lower (a difference of about 0.003), but all waveforms can be considered a match.
-xyz-grid
-Top: emonavi, Bottom: AdamW
-There's only a slight difference. Regarding the eyes, which one is closer to the original image?
-Personally, I'd say they're almost identical.
-Main Parameters
-LR: 2e-4 (50% layer-wise = 1e-4), full LoRA (c3lier), R32a16, 60 epochs/3060 steps, resolution: 1024px
-cosine with restart (3 times), bf16, Huber-SNR: 0.1
-LoRA training model: SDXL1.0 (official) vanilla, LoRA applied model: ebara_pony_3
-<Test results>
-emonavi, AdamW, can be said to be almost the same.
-Q: Does it remember details and fine points? A: It's equivalent to AdamW.
-Q: Does it cause color bleeding or color loss? A: It's equivalent to AdamW.
-Q: Does it stop remembering when saturated? A: It's equivalent to AdamW.
-So, there's no reason to use emonavi, right?
-Not at all. It eliminates the need for cumbersome hyperparameters like those for additional training.
-It automatically activates a self-repair function when there's a tendency towards overfitting or divergence.
-It converges independently of the scheduler (though using one is fine/feel free to).
-Also, special "synchronization" in unique training environments is unnecessary.
-What are emonavi's drawbacks?
-Yes, there are! Compared to AdamW, SDXL's LoRA consumes about 600MB more VRAM.
-For large-scale training, VRAM consumption for "Shadow" should be taken into account.
-Execution speed is nearly the same, but a considerable difference might emerge at larger scales.
-However, these drawbacks are minor considering the benefits mentioned above.
-＜Final Conclusion＞
-You can choose emonavi, or not—everyone is free to decide!
-<Acknowledgement/License>
-Zundamon is an official character of Tohoku Zunko and Zundamon.
-The following are the usage guidelines.
-https://zunko.jp/guideline.html
-Please refrain from commercial use.