Tan
commited on
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Wan2.2 Pixel Animate Adapter
|
| 2 |
+
|
| 3 |
+
A LoRA adapter for Wan 2.2 I2V (Image-to-Video) model, fine-tuned specifically for generating pixel art sprite animations from static images.
|
| 4 |
+
|
| 5 |
+
## Model Details
|
| 6 |
+
|
| 7 |
+
| Property | Value |
|
| 8 |
+
|----------|-------|
|
| 9 |
+
| Base Model | Wan2.2-I2V-A14B (14B parameters) |
|
| 10 |
+
| Adapter Type | LoRA |
|
| 11 |
+
| LoRA Rank | 256 |
|
| 12 |
+
| Precision | bfloat16 |
|
| 13 |
+
| File Size | ~2.3 GB |
|
| 14 |
+
|
| 15 |
+
## Training
|
| 16 |
+
|
| 17 |
+
- **Epochs**: 100
|
| 18 |
+
- **Optimizer**: AdamW (lr=2e-5, betas=[0.9, 0.99])
|
| 19 |
+
- **Gradient Accumulation**: 4 steps
|
| 20 |
+
- **Activation Checkpointing**: Unsloth
|
| 21 |
+
|
| 22 |
+
## Dataset
|
| 23 |
+
|
| 24 |
+
Trained on **226 pixel art sprite animation videos** covering:
|
| 25 |
+
|
| 26 |
+
- **Character Sprites**: Cowboys, zombies, skeletons, dark elves, fantasy characters, Santa, city characters, anime warriors
|
| 27 |
+
- **Magic Effects**: Projectiles, elemental spells, energy bursts
|
| 28 |
+
- **VFX**: Explosions, smoke effects, dust clouds
|
| 29 |
+
- **Actions**: Attack cycles (slashing, shooting, casting), idle animations, walking cycles
|
| 30 |
+
|
| 31 |
+
**Resolution**: 600x370 pixels
|
| 32 |
+
**Frame Buckets**: 8, 16, 24, 32 frames (up to 2 seconds at 16fps)
|
| 33 |
+
|
| 34 |
+
## Usage
|
| 35 |
+
|
| 36 |
+
This LoRA is designed for **Image-to-Video generation** - transforming static pixel art characters into animated sprite sequences.
|
| 37 |
+
|
| 38 |
+
### ComfyUI Workflow
|
| 39 |
+
|
| 40 |
+
Load using `LoraLoaderModelOnly` node with the Wan 2.2 I2V model:
|
| 41 |
+
|
| 42 |
+
1. Load base model: `wan2.2_i2v_high_noise_14B_fp16.safetensors` or `wan2.2_i2v_low_noise_14B_fp16.safetensors`
|
| 43 |
+
2. Apply this LoRA adapter with strength 1.0
|
| 44 |
+
3. Use with `PainterI2V` node for image-to-video conditioning
|
| 45 |
+
4. Recommended: Use with 4-step distillation LoRA for faster inference
|
| 46 |
+
|
| 47 |
+
See included `wan2-2-video.json` workflow file for a complete setup.
|
| 48 |
+
|
| 49 |
+
### Recommended Settings
|
| 50 |
+
|
| 51 |
+
- **Sampler**: DDIM
|
| 52 |
+
- **Steps**: 4 (with distillation LoRA) or higher without
|
| 53 |
+
- **ModelSamplingSD3 Shift**: 5.0
|
| 54 |
+
- **Frame Count**: 45 frames
|
| 55 |
+
- **CFG Scale**: 1.1
|
| 56 |
+
|
| 57 |
+
## Files
|
| 58 |
+
|
| 59 |
+
- `wan2.2_animate_adapter_model.safetensors` - The LoRA adapter weights
|
| 60 |
+
- `wan2-2-video.json` - ComfyUI workflow for using this model
|
| 61 |
+
|
| 62 |
+
## License
|
| 63 |
+
|
| 64 |
+
Please refer to the Wan 2.2 model license for usage terms.
|