This is an improved version of
wan2.2-i2v-A14B-Uncensored-base.
The base model has been changed to KJ FP8, and an improved lightx2v (tuned specifically for this version) has been merged into it.
This version not only works great with I2V but also
performs very well in text-to-video style. Use a blank image or the
canvas.png provided in the
demo folder.
As a rule of thumb, and to preserve the flexibility of the model, the LoRAs are merged at a strength of the “sweet spot” - 20%.
The model responds very well when used with style LoRAs or any other wan2.2 I2V and T2V LoRAs, without overpowering the base model.
If you use your own
LoRA stacks with this model, try
strengths in the range of
0.5–0.65.
The model already understands the majority of explicit prompts and the mechanics behind them.
Using
NSFW-Wan-UMT5-XXL-V2 will improve it further.
Rarely needed in I2V scenes, but in T2V scenes they can enhance appearances if the visual results otherwise look underfit.
