Goal is ultimate weird horror lora that who loves this genre!
If you have spare time to comment / feedback about this lora, i would appreciated if you do. Therefore i could work on improvements or i can switch something entirely new. Maybe mix of them. :)
do NOT MERGE it without MY permission and share your checkpoint, please!
I spent around 10 to 15 hours for each update and optimize it. I'm not throwing away randomly 10 images and clicking 1 button then release !
If you like what i do, I appreciate it if you support and buy me coffee!
Thank you!
v.1.5.2.9 has released.
Useful prompts;
Centipede
many limps
many eyes
humanoid
indoors
outdoors
blood
dismembered
dismembered body,
pov
trash
messy
Alt edition;
This is not a downgrade! It is a supplementary update to help better recognition for few " concepts".
I didn't merge them with training it is because i had to handle extra 150+ images and their captioning by hand. But i found out as a supplementary addition, it gives more details while working together if you add them seperately.
If you want to make thing works when you use both of them, make sure when you add "reelalthorror" tag, use those concepts to get better results,
eyes
extra eyes
blood
tentacles
horn
flesh
veins
multiple legs ( this might not work, sometimes it is hard to get )
ribs
teeth
open mouth
mouth
multiple eyes
Those keywords, will give you better results after the supplementary lora tag and trigger word.
Description
This is 1st version, it will probably get an update in the future.
FAQ
Comments (11)
this mod....is reached a hyper level of disgusting and scary.can't find a more accurate word to describe this.....in chinese we called this"邪门"
Hello, i liked this horror style and i wanted to share it. It was my goal to reach feel that fear tone when you see these images. I hope you find it good if you like this style. If you have any suggestion, please do it, so i can improve it, if that is possible. :) . Thank you for your comment.
Yeah...mission accomplished with this one.
Could you share your Lora training settings? this is a great model, just the right balance of style & realism, thanks.
Certainly! Training batch size is "2", Epoch = 10, Training steps = 4 ( it means in practice folder prefix is "4" ) Network Rank dimension = 256, Network Rank Alpha = 256, Schedular is "polynomial" . The reason is polynomial comes from my python experience and algorithmic learning curve. It can be selected cosine. Difference might be the same or the entirely different. The reason for it, in my experience, when you choose a style, you have to observe if it fits for learning curve. You might have that by intiution. There is no specific way to tell you at my end. For that i'm sorry :). Max resolution = 768x768. Hope it works for you! If you have any question, please do! So i can improve what i do. :)
@Reelai thank you, what platform are you using? Kohya, webui? if you could upload the json config file to the files section it would be great, I can learn from that.
@jonwong when i have spare time, i will upload when i am able to. Training program is: kohya_ss webui, standard kohya/lora learning tab. Other settings are entirely default. I didn't absolutely change anything what wrote here. This weird looking realistic produce is depending on most likely the source images, i belive. Cheers!
Hey there, sry for the delay. If i'm not mistaken, if you still have the lora file which is reelmech, you could able to click on the "i" icon if you're usin webui and see all the parameters that you're looking for.
can't wait to try 😁👍
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.