Pixel style, side view attack animation.
Please download the full lora with high-noise and low-noise from this link:theamusing/wan2.2_pixel_attack_lora at main
I have uploaded three checkpoints with different training epochs. The model is trained on both high noise and low noise loras, so you need to download pair of them.
Training
Trained on around 30 melee attack sprite animations with resolution 512*512 and 28 frames.
To use this lora
Input images should be low-resolution (around 64×64) pixel-style characters resized to 512×512.
Since all training data face to the right, it’s recommended to keep the input character oriented that way.
You can combine this LoRA with 4-step LoRAs such as lightx2v for better results.
From my experiments, higher-epoch checkpoints tend to consistent slash attack, while the low-epoch checkpoints make the generated actions more aligned with the input.
High-quality pixel-animation resources are quite limited. Currently, there are no stabbing attacks in the training set, so the LoRA is more likely to generate slashing attacks.
It’s still hard for this LoRA to produce controllable multi-hit combos 😅 That’s on the to-do list.
Description
FAQ
Comments (8)
Truly wonderful work thank you!
I see the HF link, but what is the file you have attached here?
If you mean the file here on civitai, it's a high-noise lora checkpoint for 1750 epochs. It's just to meet platform requirement. Please use the models on HF.
If I may ask : What does the number of epoch change in the output video? Thank you
From my experiments, higher-epoch checkpoints tend to consistent slash attack, while the low-epoch checkpoints make the generated actions more aligned with the input. The example image are generated using 1250 epoch model
@aaaa_Aa Thank you !! :)