EDIT : V1.0 was trained and turned out way worse than V0.5, and I don't feel like trying to figure out why and train again, because this was just a test model anyway, so I guess V0.5 is the 'Best'... I still decided to upload it, but honestly it's not good, feel free to test, maybe you'll find a good balance and way to use it (I didn't...) I will upload the prompts used in training as well for those who want to experiment from them. for what it's worth, try V1.0 not higher than 0.85 it may be overtrained (and that may be part of the problem)
V0.5 Gamma version TIPS : Describe nuclear explosion in the prompt and be creative (or destructive...) trained on clips from real nuclear explosions, from various tests, most of the clips are from 1945-1956 around that, very few in full color (often orange, yellowish, and black & white, one or two are in full color, faded colors mostly, take this into acount if you want to counter this by prompting something else of course.
Early version, not the real model just a preview (I wanted to be sure a training will yeld something over 5 kilotons) I'll add some deuterium and few copper pipes in the second version, salt and pepper and see where it goes, few laser to compress the core and we're good for the big Boom.
This test training session was done with only 9 kilotons clips (final version might have as much as 40 megatons, but not yet sure if I will go that far) and be able to also depict shockwave blasting effect and total destruction (at least in theory, or it will just destroy the galaxy if it goes really really bad) this training was done with ZERO prompt, meaning I've let the model figure it out (which is certainly not optimal, but I will add prompts in the second version) Only real nuclear explosion clips used, no fake, no video games, this lora is radioactive at this point. Put your protective glasses on, don't be closer than 45 miles before each video generation, only generate ONE video per month if not you will exceed deadly dose of radiation (more than 500000 Grays per video at impact point)
Probably too much diverse clip killed the flow of the model, it is in theory better and more able to capture certain aspect, but seems very bad at actually doing it... still, here for anyone wanting to test and maybe figure it out better than me... I won't do V2.0 no time to train this one again, I'll focus on other stuff
Go ahead and upload yours!
Your query returned no results β please try removing some filters or trying a different term.