Cacao | Nekopara

a LoRA model based off Cacao from the Nekopara anime adaptation

6.1 MODEL RELEASE DATE: Never? (By me atleast)
Whoever would like to create Cacao 6.1 is happy to do so, full permission granted. Just make sure to ping me so I can see the full product myself :3
Here is the link for the zip file as I've abandoned this LoRA and LoRA creation in general:
Link

Very late on this but I thought I give updates on the why:

• When I was making 4.0 Google Colab was free to use with Stable Diffusion, 4.0 became a version that is usable for the most case but only for creating Cacao, so there wasn't many variety, and generating those were painfully hard.

• 6.1 consisted of 287 images, literally every scene where Cacao appeared, or was in a different outfit got a screenshot, still image. I literally rewatched the entire Nekopara anime countless of times

• Since Google Colab became a paid subscription I decided to give one try with the free one, it took hours and then obviously the gen stopped.

• With years passing by I lost motivation, so much so that I eventually deleted the zipfile of Cacao 6.1 LoRA and altough it might be available somewhere in bits, I would likely have to start all over again with prompts (but If I were to find it I promise I upload it as a zip file for anyone who knows how to create LoRAs correctly, as I was looking for tutorials even while making 4.0 so I was happy that so many people like it despite being under or overtrained.

• For me to create 6.1 would recquire to either start over or have the patience and mindset to go through with this which honestly I don't

• Last but not least, the model was suppose to be an SD2.5, and I don't know how relevant it is to use SD2.5 nowadays with XL, 3.5 being a thing.

Here's what wouldve been new in version v6.1

- Trained on 237 images instead of the previous 81

- Extreme focus on generating high-quality images

- Every outfit from the Nekopara anime is now suitable

- Every pose that Cacao had in the anime now trained in the LoRA

- Better prompting, focus on facial expression, variety enhancements

- 768x768 picture training was used instead of 512x512

- better focus on backgrounds

- Environments are now a key factor (meaning: objects, wind, indoors, outdoors working better)

- Fixed and minimized overtrained and undertrained images

- Captured the Nekopara art-style via training the Lora

- It is both trained on the AnyLora and AnythingElse V4 checkpoint models

additional information, generation guide will replace this textbox on release, this is just a tease on what to expect. Keep calm :3