Art style for backgrounds from Westwood Studios' The Legend of Kyrandia trilogy.
This was quite an instructive case for an amateur like me on how differently Anima trains compared to Illustrious, especially when it comes to preserving a pixel art style.
Trigger word: kyrandia-bg
Additional reinforcing tags: pixel art,dithered background,dithering
V2-epoch9: Westwood art style flourishes with noticeable Anima detailing.
V2-epoch6: My favorite from the base Anima samples. Balances the pixel art style and shading with some coherent Anima polish.
V2-epoch4: Blockier and more pixelated - possibly easiest to retain the blocky pixel art style.
V1: Default training settings - art style way oversmoothed by Anima's training tendencies.
Description
Trained on ~150 images from Kyrandia 1 and 2. Tag-based training data annotation.
FAQ
Comments (5)
Overall, it’s very good in terms of colors and general style, but it lacks large pixels—though that can be tweaked using other LoRAs. Thanks anyway! :)
I played around with the training params and uploaded three epochs from a V2 training that demonstrate the tug of war between the pixel art and Anima's desire to smoothen.
Seems to be getting closer to the in-game style, though it still struggles to match how well SDXL architectures adhere. Let me know what you think of the new ones.
@phenoxide It’s strange—one of the three new versions failed to install entirely, while the two that did install yield poorer results than the first version. Furthermore, the pixelation is actually worse when NOT using a LoRA... Is it normal for these new LoRA files to be 17 MB, compared to the 130 MB of the first version? I highly recommend reaching out to the creator of this LoRA to ask about their training settings, as it works perfectly. https://civitai.red/models/705520/heroes-of-might-and-magic-2-style?modelVersionId=3411407
In any case, thank you very much for your efforts; the first version still produces good results.
@laska_ai Thanks for reviewing and for the feedback. Anima (and DiT training in general) is new to me, so I appreciate the comments.
(1) Strange that one failed to load for you - do you recall which? I tested all four with the onsite generator shortly after posting and had no issues.
(2) When you say pixelation is worse without the LoRA, do you mean the same prompt becomes over-pixelated or under-pixelated for the Kyrandia style? To my eye v2-epoch6 got closest to matching the visual language of the games, but still falls well short of Illustrious.
(3) The file sizes are expected. The V2 epochs were generated with much lower network rank and alpha settings, which results in a smaller and less complex LoRA. I did this to try constraining Anima’s tendency toward smoothness and fine-detailing, though I suspect that I overshot the sweet spot.
(4) The HoMM2 LoRA is beautiful and very faithful to a lightly upscaled take on the original portraits.
@phenoxide I think the issue with loading one of the files stems from the fact that I work exclusively using Invoke—which, unfortunately, is no longer being updated—and I suspect that’s the root cause.
As for the pixelation, what I meant was that when using the same prompt without a LoRA, the pixels in the image appear larger and more pronounced compared to when a LoRA is applied. :)
I’m sure there are plenty of complexities and challenges involved in training your own LoRAs, so please don’t worry too much about it! :) In any case, we all have to start somewhere. :)