■This is an experimental ground for krea2 LoRAs.
■If you notice pixel patterns in your generated images, try the VAE below. It removes the pixelation typical of the qwen_vae and makes the images much sharper.
https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image
■If the VAE looks a bit too saturated for your taste, try using this VAE merge node to blend it with the original qwen_vae to dial it in. A 0.5 merge should be just right.
https://civarchive.com/models/2487550/vae-merger-for-comfyui-anima
Description
■This is a Megaman X style LoRA, trained on 400 carefully selected images from "Megaman X DiVE" and "The Day of Σ" anime.
Roughly 150 images from "Megaman X DiVE"
About 250 images from "The Day of Σ" (official FHD video)
Rather than a pure "style" LoRA, it might be more accurate to think of this as a LoRA trained specifically on the physical features of Megaman characters.
■Features & Usage
Recommended Tags: Adding tags like android, robot, robot joints, joints ,etc... will really help bring out that classic Megaman character vibe.
Visual Characteristics: It makes it easier to generate human-machine hybrid bodies. It also emphasizes a manufactured, metallic gloss on the body, and eyes tend to be drawn a bit larger. Since the DiVE sprites heavily use rim lighting, you'll probably see that effect a lot.
■A Note on Backgrounds
This might be a bit divisive, but since the Megaman X DiVE dataset consisted entirely of character sprites on white backgrounds, I used "nanobanana" to add backgrounds to the images before training. I did this because training strictly on white backgrounds makes the model prone to generating unwanted white or overly simple backgrounds.
I carefully considered what backgrounds would suit each character and used detailed AI prompts to generate them. Because of this, you might notice an "AI-generated" feel in the backgrounds, lighting, and overall style.
If the style feels too strong or unnatural, lowering the LoRA weight/strength should help reduce the effect.
FAQ
Comments (5)
as always krea is the better of the two but slow as a a snail (i use a AMD card) but as for future lora ideas.... 1.digimon adventures 1 and 2 and no more ,2kingdome harts the artwork not the 3d stuff , the reson i put theas two up as there are no good style loras for krea only charater ones.
Thanks! I feel the same about Krea 2's speed; even the turbo model is pushing the limits of my patience. Doing large-scale training locally is just becoming unrealistic.
Thanks for the LoRA ideas! I lived through that era but surprisingly missed out on those series, so they sound interesting. I might try them out once I finish my current projects.
■Since there are so many reference materials for Digimon, the style training itself should be pretty easy.
So, if the goal is purely just to capture the art style, I think I could whip up a quick style LoRA using a simplified dataset without spending too much time on it.
As for whether I make a "v02" that covers specific characters and the entire story, I can just leave that for later.
Even then, with so many episodes, it would be really tough to cover every single character and story arc, so I'd likely have to compromise somewhere.
It could easily turn into a massive, 1-2year-long project, so it really just depends on how much passion I have for it! lol. I might give it a shot if people actually want it...
■Kingdom Hearts is definitely a challenge. It's the opposite problem—I'm worried about the lack of available data and the resulting quality. Still, plenty of people make amazing LoRAs with small datasets, so I'll try to be optimistic!
This variant is much smaller and less RAM-hungry and does not slow my pc down when gening https://civitai.red/models/2805755/project-chimera?modelVersionId=3194756