This is a character from a cover illustration by Angus McBride for a History textbook.
Although the details are unclear, I believe this illustration was used in some board game at the time.
Include these prompts at a minimum :
muscular, blonde hair, dark blue eyes, facial hair, beard, manly, barbarian,
Clothes(?):
helmet, holding, holding polearm, polearm, sword, shield, weapon,
Since Gemini has generated the “penis”, unless otherwise specified, a penis will appear and you will be prompted to pay the yellow BUZZ.🙄
If you use the trigger word “gaesata,” the art style will be closer to the original artwork. If you use the trigger word “anisata,” the art style will have an anime-like feel.
In the original artwork, he was a rough-looking, wild-man-like middle-aged guy, but the nanobanana filter has turned him into a “daddy”…
Description
Comments (8)
how did you get the dataset
I had Google's Nanobanana generate it. That's why the face looks different from the original illustration.
@shin_ramen I don't know, but it just works so well even if it's just for the historical clothing. How many images did you use and how did you do it? still, I'd recommend you to train this as a concept rather than a character
@Telekinetic_tomato Since I liked not only the costumes but also the character designs themselves, I tried using “Characters” as a model for my training, but it might be a good idea to present them as a concept as well😆
@shin_ramen well, how did you exactly do it?
@Telekinetic_tomato I’ll provide nanobanana with the original illustration and have them create full-body images of the “figure” from the front, side, back, and a diagonal front angle (4 images so far).
Next, create full-body illustrations in anime or Copic-style art, and create the same four angles mentioned above using a “cowboy shot” perspective (12 images total so far).
Then, create images facing straight ahead and looking diagonally forward from both “from above” and “from below” perspectives (16 images total so far).
Finally, create 20–30 full-body images featuring various poses and expressions.
Twenty images should be sufficient, but if you come across any that look particularly good, feel free to include those as well.
After that, feed the images into Civitai’s training system for auto-labeling, adjust the tags, and train it using 12 epochs and 20 repetitions (which amounts to about 1,200–2,000 steps). That’s it!
@Telekinetic_tomato When creating a LORA for a character with a complex design, you can prevent overfitting and stabilize the design by adding the same number of images of similar-looking characters with comparable compositions to the 20–30 training images you created this time, using them as regularization images.
In this case, do not use the trigger word feature during the training phase; instead, enter the trigger word as a tag directly on the image of the character for which you want to create the LORA.
@Telekinetic_tomato By the way, for “gaesatae,” I’ve prepared 16 to 24 training images for each of the two art styles. By tagging one set with ‘gaesata’ and the other with “anisata,” I’m testing whether the model can treat the style not included in the prompt as a regularization image during image generation.
We’re using this approach because gaesatae is a nearly nude character. While nanobanana tends to avoid generating full-body images for figures or anime-style art, it was able to generate full-body images (including male genitalia) when using the original art style.