When I realized that AI often struggles to generate good-looking images with sabers or swords (especially how characters hold them; the hand often looks weird, with extra fingers or unnatural grips), I decided to create my own LoRA.

This LoRA is primarily trained in a comic book style, because that's the kind of image I wanted to create. I haven't tested it much with other styles (like realistic), but feel free to try. I might create another version in a different style later on.

I trained it on around 100 images at a resolution of 832×1216, with a variety of poses. It works well with tags like:

• “holding a [color] lightsaber”, “holding a [color] lightsaber with both hands”, “holding two [color] lightsabers”, "comic book style", "dynamic pose", "clean line art", "bold ink lines", "cel-shaded", etc.

You can also add pose tags like “in a defensive stance” or “attacking”, etc.

Of course, it still struggles to generate exactly what you want every time, but if you create a batch of 5 or 6 images, you'll usually get at least one that works. I personally find that a weight of 0.5 gives the best balance (higher values tend to make the output too "rigid").

Tip: I highly recommend using inpaint around the lightsaber and hand area. Even if the first generation looks a bit off, reworking just that zone usually gives excellent results, especially for correcting fingers or improving saber grip.

Recommended usage:
• Style: comic book, cel-shaded, bold ink lines

• LoRA weight: 0.5

• Base model used: DreamShaper