2000 steps trained lora for IllustriousXL+Trex style 2.5d generations. SFW & NSFW.
Trigger: 2.5d
Run at strength 1.0 for optimal results. The lora works also well anywhere withing 0.7-1.0. Tested on Euler/simple on both Fastvideo H3 and the usual Turbo loras, so you can run it on 6-8 steps no problem. Best results with 20 steps without turbo lora (marginal improvement for those who want maximum quality). I trained this lora also all the way to 4000 steps but that overbaked it so I retracted back to 2k steps which provided a much more consistent result without overcooking the style completely.
The lora is trained for Ref2va, but works also on first-last-frame.
I wanted to create this lora because Minimax defaults to a pure 2d style (anime) while struggling to maintain this specific style if using ref2va. I got tired of generating first or last frame images so I decided to create a proper style lora.
I will work on creating more flexibility with body parts in the next version of this lora, possibly merging it with some NSFW loras for actions.
Let me know in the comments how the lora worked for you!
-Stuubzzz
Description
FAQ
Comments (4)
Some parts became blurry or smeared, so the result wasn't very good. I used the same sampling method, but the generated video still didn't turn out well. It might be an issue with the workflow. If I could take a look at the workflow, I might be able to figure out where something went wrong.
I'm currently training v2 for this lora. There are some checkpoints or models this lora doesn't work with that great (singularity, 10eros to name couple). If you want to see my workflow, it's this:
https://youtu.be/aYeyDqsjpvM?si=kEi3Dx-PAB74LdTd
@Stuubzzz I hope that for future LoRA releases, you could include the test prompts, sampling methods, and the workflow used to generate the example images. This would make it easier to evaluate the LoRA's actual performance and also understand how you use prompts with the model. Since this is a paid model and there aren't many examples available, having this information would be especially helpful. Thank you!
@lq122449105 Hi! Feedback taken. I will improve for my next LORAs and add those!