license: apache-2.0
base_model:
- Wan-AI/Wan2.2-I2V-A14B
This model is suitable for use as a base for merging. | LightX2V 4Step LoRA
I collected around 200 highly engaged prompts from various communities and analyzed their common patterns, structures, and phrasing. I then combined this analysis with the official Wan2.2 guidelines to create a specialized system prompt for JoyCaption.
This system prompt was used to caption the training dataset in a style that matches popular user prompts while still respecting the model's official guidelines. After 4 training sessions, the resulting LoRA became quite reliable.
Strengths:
- Understands short, simple, and uncensored prompts out of the box
- Solid male & female anatomy (7/10 reliability)
- Follows user prompts closely
- Excellent flexible base for merging other LoRAs
- Adding LoRAs rarely overpowers the model
Recommended LoRA:
CubeyAI-GeneralN-High.safetensors & CubeyAI-GeneralN-Low.safetensors
Using these LoRAs with the official Wan2.2 barely does anything, but when paired with this merge at 0.5 or 0.55 strength, it improves fluid and rhythmical motions and often remove the need for other LoRAs.
Compared to official Wan2.2:
- Much better prompt understanding
- Good basic mechanics for most actions
- More stable when adding LoRAs
- Low strength LoRAs usually improve quality
Limitations:
If the input image doesn't contain the objects or people you want to add or modify, the model may still struggle, even with LoRAs.
Workaround: Place a brightly colored distinct shape in the scene and instruct the model to replace it. This works very well with first/last frame techniques for face swaps and continuity.
Be descriptive with actions and objects that are not visible in the input image. UM_T5_XXL is not a strong reasoning model, so vague prompts often fail when adding new elements.
Example style:
"A {fully/partially nude} young man with erect *** {approaches/appears/enters the frame from the (left/right/behind) }, her, and thrusts his **** back and forth in ***..."
Sound:
- MMAudio --> Best for background & sound effects
- SkyReel V3 --> Excellent for talking
- LTX-2.3 --> Fast but less reliable (better on >7s videos)
Related LoRA:s