Tag word is: Remilia Scarlet (Her full name)
A Flux Dev Lora of Remilia Scarlet, much better trained than my Shamiko lora. Works best at 1.0 weight, can be turned up slightly down or up depending on what you need.
As always, no trigger word, describing her appearance can help immensely, short blue hair, red eyes, mob cap, pastel pink outfit, dress, etc. The lora used a diverse dataset so other outfits and styles are possible.
Training info:
33 images, 3000 steps, Batch size of one, at 1024x1024 resolution, the learning rate was set to 1e-4, I used the simpletuner training script and trained on 1x3090 for 3 hours. Other experiments, like combining this Lora with Sakuya were tried on my other two 3090's but did not turn out.
I don't like buzz, don't give me any please. I'd much rather you leave a text review or share some generations! <3
Description
Initial upload
FAQ
Comments (10)
Great lora! Can you share your config.env and multidatabackend.json?
@Zennou Thanks!
Easily my favorite Lora for Flux so far! Gonna have fun with this one!
How was your dataset captioned?
Natural captioning using GPT-4o and Claude. I'm currently experimenting with other methods.
Hm. You used deepbooru tags right? And the dataset only consists of 33 images. Will you ever do the training with larger batch size?
Nope. I used natural captioning for this model. I am looking into other methods right now, tweaking numbers, batch sizes amount of images, etc. For simple characters like Shamiko or Remilia it seemed to work fine, but more complex characters like Flandre Scarlet seem to need more data and epochs.
Only problem I've been having is with her wings. Could just be a FLUX issue as a whole but it very rarely attaches her wings correctly if you do any angle that isn't from the front. Other than that it has been great so far.
nice work! any other touhou project character please~
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.