It creates from cold dark images to warm golden light the negative values can be a little bigger than the positive ones. From -6 to 3 but it changes for each image
Description
First version from -6 to 3
FAQ
Comments (13)
THANK YOU SO SO SO SO MUCH
Thanks for doing my request! For some reason, I got you and loraholic mixed up as I requested him to do one. Consider me drunk lmao
你做了测评吗，这两位创作者的区域高度重合，我实在是对比不出来效果
Thank you!!!
really useful for correct lora light bias
All the loras you've shared are really helpful, my friend. Thanks!
Do you happen to have a guide or tutorial on how you are making the Krea2 loras? I have a lot of concepts I'd like to try making sliders for and would love to know what your best practices are and what tool you are using?(Especially the parameters used)
Unfortunately I don't have. I'm sorry. But Youtube is full of great tutorials and with a little effort you be doing it. Ostris has a great tutorial for ai-toolkit sliders.
@alcaitiff Thank you for the recommendation, so are all the sliders you made using iToolkit?
@lse14 I modified my ai-toolkit version to change some things, one thing comes to the other and now I have a personal fork. I do a lot of weird manual surgery on my LoRAs: changing ranks, pruning layers, tweaking parameters in the middle of training, running ablations, checking the impact of individual layers one by one, and generally poking the network with a stick to see what breaks. That kind of workflow ended up needing a bunch of extra features that didn't really exist in the original toolkit. So no, my loras are a mix of all things I learned and created in last years.
@alcaitiff No wonder your slider Lora effect is so good!
In addition, do you have any good suggestions for using rank, caption, and training set for the basic aitoolkit training slider?
Sliders can work perfectly fine at rank 1. Higher ranks add capacity, but they also make it easier to learn things you didn't want. Check the tensor view on Civitai to see the rank of LoRAs you like, you'll find plenty of good low-rank examples.
The important part starts after training. Test your slider at extreme values, not just ±1. That's where defects show up. A hand-size slider, for example, shouldn't change lighting, colors, noise levels, composition, or character gender. If it does, something leaked into the concept.
To fix it, either improve your training prompts or inspect the layers and prune/scale the ones causing side effects. This is where a lot of the quality comes from.
The good news is that sliders train very fast. Don't treat them like huge LoRAs. Run lots of experiments, break things, retrain, prune, compare results, and repeat. You'll learn more from ten bad sliders than from reading ten tutorials.
@alcaitiff Your experience has been very helpful to me. Thank you very much!