I love Graf Zeppelin, so I trained this model and use to collate and write the LoRA training document. You can also check https://github.com/28598519a/lora-scripts-windows.
I think the most difficult thing of train this LoRA is to have a good finger, iron cross, the symbols on the capelet at the same time.
I've uploaded the preprocessed training data (located in the v1.0 option), thanks to the artists.
Recommend to use Anythingv4.5 to generate anime-style images (AnythingV4.5 > AOM2 > AOM3 > AnythingV3/V5、Counterfeit > NAI)
Usage:
You can choose
graf zeppelinorgraf zeppelin \(kancolle\)as the main trigger word, but don’t use both (graf zeppelinis recommended)The main style of Graf Zeppelin was split in 4 tags (
uniform, capelet, hat, machinery), usually not usedmachinery(艤装 / ぎそう).You can select some keywords to add in prompt e.g. (
swimsuit, bikini, playboy bunny, maid, nun, japanese clothes, christmas, abyssal).If you used v1.1 on a realistic model like chiloutmix-ni, the weight should give higher like <lora:grafZeppelin_v11:1.2>
Description
Finger generation doesn't perform well, but I share it because when ignore the finger part, it usually produces good illustrations.
FAQ
Comments (7)
Well I have to clarify that you are not properly using my model for comparision, that's a bit unfair. The trigger word is totally different. I mean yeh u can make a compare but can you read the introduction before using it?
Your trigger word is graf zeppelin \(kancolle\), so the comparison table by your model is all use graf zeppelin \(kancolle\) to compare with mine, please check the left side text of image. Also for realistic comparison, I use both graf zeppelin \(kancolle\) for prompt yours and mine. (This provides a fair comparison of the basic types of character.)
And the outfit you claim are all graf zeppelin character's original element, and my table proves the word "has a wide variety of possible clothings" in your introduction doesn't seem to be the case.
Based on the previous message, I think this is a fair comparision.
If you have any ideas for comparison, you are welcome to provide one too.
@RipTheDarkness Not really. I didn't remove any tag during training and didn't use keep prompt to train it, that's why I have mentioned that if you want to generate the clothes, you should input all the prompts I gave, About the variety of clothing, yeh, it's a bit overfit, but not impossible.
Anyway, I do have noticed that my version is bad when the prompt isn't properly typed and combined with using highly mixed models (AnyV4 and AOM3) though. But still, I won't say it's a fair comparison. Maybe will update one sooner or later.
@4BEraser It seems that you leave the message without reviewing the training data I provided, this version I didn't remove any tag and didn't use keep prompt also.
And again, about the comparison, you are welcome to provide one too.
(And just a friendly reminder, your models problem is about overfitting, I suggest you to thinks about your training data and training parameters problems and do some test.)
@RipTheDarkness yep, I have to admit that, but still less weight and proper negative prompt will help. Will show you soon. And I have found that your model isn't working well on AnyV3 either. Do you train on Any v4.5?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.