Trained on screenshots/generated images from the ARPG game ANIIMO.
✨ If you enjoy my LoRAs and creations, you can support me with a little coffee on Ko-fi ☕💖
It helps me keep training models, creating art and making more dreamy weird little projects 🌙✨ Thank you so much for the support and kindness ~
Description
Strength/Weight: 0.8 - 1.0
Trained on 14 screenshots/generated images of fenmane ~
FAQ
Comments (4)
can't wait to give these a try. I tried to do the same thing, make a few LoRA's from Screenshots but it came out heavy 3d no matter what.
mind if I asked how you did it?
also, what do you mean by trained on "generated" art? how could you train on generated art if there was no real way to generate them without a proper working LoRA?
also, Generated Images used as training data can be more of a negative than a positive as details are never exactly correct so it can cause character Inconsistencies
@Tundra1996
Hi, Tundra ~
When I say “generated” I didn’t really mean raw AI outputs thrown straight into training, sorry. I meant reconstructed base images that were corrected and refined before ever touching the dataset.
My process for this was:
• I cleaned and reconstructed the images first (background removal, artifact cleanup, light normalization when needed).
• Some images were then partially or fully reconstructed through generation; sometimes with completely different angles/poses not in the screenshots (using ChatGpt/Flux Klein with multiple reference option - lots of hit and misses) and then manually corrected/redrawn in parts to match the character pretty much 1:1. If something feels slightly off in structure or color I usually try to fix it, especially with smaller datasets (easier and less time consuming).
• Lightly upscaled and softly sharpened each image individually (almost never in bulk). I adjust per image depending on what it actually needs.
• Heavily curated tags (adding "3d" especially helps)
For Witchin as an example, I generated a few cleaner 2D-style reconstructions (converted from 3D; also had to erase quite a few hands since she has none and then draw some tentacles instead) and tagged them accordingly with '2d' so the model doesn’t cling to the engine look. With fenmane I didn't use any 2d style converted images though so the 3d does push through more for that character lora.
• Afterwards when generating images (like the preview images) I will use either no lora's or several lora's and have "2d" at a higher weight in the prompts and "3d, cg, realistic, live action" in the negative prompts, for example.
Generated images can absolutely cause inconsistencies especially when left on their own.. but when corrected/standardized first they can help reinforce missing angles and poses that raw screenshots simply don’t provide at times.
Hope this helps! (/ω＼)