Usage of v1.0 and v1.1
Trigger words: She is lifting her tops, revealing her breasts and nipples Titty drop, flashing her breasts
Works with fp8, gguf variants, and FastVideo
LoRA strength: 0.5 - 0.8
Guidance scale: 9.0
Prompt: see the examples
Training details
See the "Training Data" file
Usage of v0 (not recommended for new users)
Trigger word: t1ttydr0p
Works with fp8, gguf variants, and FastVideo
This is still experimental and it seems not activated simply by the trigger word. You need to experiment with seeds and prompts to get the right result. Here is an example settings that I found it relatively working (also see the example videos and workflows):
num_frames: 41-61
seed: 665062624858284 (use this as a starting point)
Prompt:
t1ttydr0p, A professional video of a skinny Korean k-pop idol girl dancing on the concert stage. She lifts up her crop top shirt. After a second, she reveals her small breasts and nipples. She wears a crop top and a pleated mini skirt. In the background, there are other girls in the same group fully dressed. navel. Detailed smiling face with gorgeous makeup, high resolution.Sometimes, you need to reverse the video. I am not sure how to fix the training process to solve this issue...
About the example workflows of v0
HunyuanVideoWrapper (faster, recommended): Optimized for RTX 4090 or above (24GB+ VRAM) with sageattention and triton
Comfy: Uses a gguf and the official nodes (optionally you can use sageattention for faster generation)
Description
v1.0 comes with significant improvements:
Adhesion to the prompt got significantly better
No need to find good seed values anymore
Trigger words are changed to "Titty drop, flashing her breasts"
Notes
Recommended LoRA strength is 0.8. You will lose details if it is set higher.
The training dataset does not include many videos with full face. You may need to describe her face in the prompt.
This LoRA also changes the "default face" a little. I will try to fix this in the next version.