Loaded
Download
1 variant available
bf16 SafeTensor
_H3__Golden_Sun_Style_minimaxh3_3364911_epoch_18.safetensors
BF16, good balance • 295.8 MB
Verified: a day ago
525
4.6K
13.8K
Generation, training and LoRA distribution on Civitai are covered by Civitai’s own license agreement with MiniMax. If you download these weights and run them yourself, your use is instead governed by the MiniMax H3 Community License Agreement, whose grant excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America.
MiniMax H3
Attempts to replicate the out of battle scenes from Golden Sun and its sequel on the GBA. I have no idea if this will work, but I hope it does. Since I have little experience with natural language tagging (and would love feedback on it), there’s some small concepts that appear a few times that aren’t worth listing the training language for on their own, and I can, I have made public a zip containing the txt files with the tags used so people can review them (it's under optional files).
Click here to see 70+ LoRAs otherwise hidden from my profile at the whims of credit card companies. You can help support making models like this by posting stuff you've made with it to the gallery using the "add post" button, that helps me earn back the training cost at no cost to you.
Trigger: A Golden Sun pixel art style video (or image or character sheet)
Made with 19 clips of video footage (the longest is a 32 second piece of Yepp’s Song, and the second longest the water bearer puzzle in Xian at 29 seconds, but most are under 10 seconds) from the two games plus 10 stills and 6 sprite sheets from the two games, all hand captioned by me (did throw a few images at auto-label to see what it was looking for, but most results were garbage and I only wound up describing certain things it in my own words after seeing those particular features caught the auto-label’s “eye”). I made some effort to train text boxes (sound effects that play for text including) and background music (though I don’t really know enough about music to do more than say its there). I wonder how those turn out?
V2 adds 15 new videos, bringing the training data total to 50, and making the longest a 38 second scene in Kandorean Temple and second longest a 37 scene in Madra (in general these new clips are longer, most over the 10 second mark and several over 20 seconds). It also features small changes to existing tags, the biggest being to identify Felix and Jenna as characters (I think it will allow them to appear more in the dataset without making everyone look like them, and may make them promptable, but I don’t expect that too perform too well)
“Felix (A man with a blue cloak, ahoge, low tied brown ponytail, ahoge, brown gloves, and boots)”
“Jenna (a young woman with red hair in a ponytail, a blue cloak, a shirt, a skirt, and boots)”
Since H3 seemed to know lots of cartoon styles well enough for me to not feel it was worth bothering with teaching them, I figured I’d try video game clips and see what happened since it also freed me of needing to trim videos into neat clips since I can just record the clips as clips. Big thanks to golden sun syndicate user “Atrius“ for a nearly 20 year old set of GS codes for changing the map sprite of the player and to archive dot org for preserving them: Made it way easier to get footage without the main character being in every video.