Penguin Empire Mooks - Kenzen Robo Daimidaler (optimizer tests)版本v1.0 AdamW (ID: 101369)

MAE SHIPPO! IT'S a FRONT TAIL YOU DEVIANTS!

V2:

This time I tested the Prodigy optimizer, same dataset.

Prodigy LR=1 DR=.01 Dim=32 Alpha=16

Vs

AdamW LR=.0001 DR=.1 Dim=32 Alpha=16 Scale Weights Normals

vs

AdamW LR=.0001 DR=.1 Dim=32 Alpha=16

The result: Contrary to DAdaptAdam I got good results with Prodigy. In good conditions its behavior is similar to AdamW but with 15% more cost in training time. BUT! In less than ideal conditions when you are close to overbaking and using extra images to fight against bleedover, it performs significant better than AdamW with Scale Weights Normals and STUPIDLY BETTER than stock AdamW.

Conclusion: I would indeed recommend Prodigy for beginners as it is much more tolerant to overbaking and imbalances in the Learning rate vs repetitions vs dataset size. It also seems to help a bit with concept separation(but I tweaked everything as much as possible to maximize it in my sailor moon LORA.). The only issue is the extra required configuration it needs(It doesn't need to be adjusted). I added a pre configured template in my training guide if someone wants to check it out.

V1:

I needed a sacrifice to test the DAdaptAdam optimizer. Thus I trained two loras one as normal with

AdamW LR=.0001 DR=.1 Dim=32 Alpha=16

VS

DAdaptAdam LR= 1 DR=1 Dim=32 Alpha=16 25% slower

The result: The DAdaptAdam sucked so much ass compared to the normal recipe despite me using a default not fine tuned learning rate. Maybe it will get better with further experimentation but right now doesn't seem like it's its moment to shine. The fact it took longer makes it even worse.

Anyway...

As I didn't want to sacrifice a pretty lady for research purposes I used the Mooks of the penguin empire of the sadly underrated Daimidaler series a hilarious parody of the super robot genre.

It's refreshing when you agree with a series point of of view, the main characters of the second part of the series are extremely fucking annoying. The leader of prince was RIGHT!

Anyway what you are here for, humanoid penguins with front tails, got ok results with:

Positive: PenguinEmpireMook, standing, solo, masterpiece, best quality, highres, cloudy sky, <lora:PenguinEmpireMookV2:.7>

Negatives: lowres, bar censor, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, censored, ((depth of field)), simple background, eyes, face, cape,

As they have no faces or definite hands, bad hands embeddings and face positive tags will likely make more damage than help. That's why I added the negatives too.

Clothing sets? I guess one guy uses a hat once.

For the dataset, well it is mostly screencaps so it is a no go, instead I will dump the training files and the other lora which I dare not publish(due to crapyness) if someone wants to review them.

描述:

训练词语: PenguinEmpireMook