Cxy Kiss H3 LoRA
A specialized video LoRA for MiniMax H3, designed to improve kissing interactions in AI-generated videos.
The goal of this project is to help MiniMax H3 generate more natural, coherent, and convincing kissing movements between two characters, with particular attention to mouth contact, movement continuity, and overall interaction quality.
This LoRA has been developed through multiple rounds of dataset refinement and training experiments.
Official Accounts
My Civitai account under the same username was also mine. This Hugging Face account is mine as well.
Both accounts belong to me.
Model Versions
V0.3
- Trained on approximately 300 videos.
- Uses a relatively large and diverse dataset.
- Focuses on learning kissing movements and interactions across different scenes and poses.
- A broader training dataset designed to improve overall motion consistency.
V0.4
- Trained on 77 carefully selected videos.
- Primarily trained at 768 resolution.
- Uses a more focused dataset with an emphasis on high-quality motion samples.
- Improved mouth detail and visual clarity compared with the previous 512-resolution training approach.
- The 500-step checkpoint provides the best results in my testing.
V0.4 demonstrates that a carefully curated dataset can be more effective than simply increasing the number of training videos.
Recommended Setup
Base Model
MiniMax H3 Hybrid FL2VA / Ref2VA
Recommended checkpoint:
minimax_h3_hybrid_fl2va_ref2va_b25-49-int8
Acceleration LoRA
MiniMax H3 Turbo V4
Recommended model:
minimax_h3_turbo_v4_step600_ema_pruned
- Recommended strength: 1.0
Kiss LoRA Settings
- Recommended strength: 0.5–1.0
- Workflow: ComfyUI with a compatible MiniMax H3 workflow.
- Ref2VA workflows are recommended.
Adjust the LoRA strength according to your prompts, reference images, and desired results.
Notes
- This is a video-generation LoRA, designed specifically for MiniMax H3.
- Results may vary depending on the base model, workflow, prompts, reference images, and sampling settings.
- Different LoRA strengths may produce different motion characteristics.
Feedback, suggestions, and prompt-sharing are always welcome!