wan22-fp16-i2v-loras

<!-- README Version: v1.4 -->

WAN v2.2 FP16 Image-to-Video LoRA Collection

Specialized LoRA adapters for the WAN (Wanniwatch) v2.2 video generation model, providing enhanced camera control, lighting effects, quality improvements, and character actions for text-to-video and image-to-video generation.

Model Description

This repository contains 9 specialized LoRA adapters designed to enhance WAN v2.2 video generation capabilities. These adapters provide fine-grained control over camera movements, lighting conditions, facial animation, and overall video quality without requiring retraining of the base model.

Key Capabilities

• Camera Control: 5 specialized camera movement LoRAs (rotation, drone shots, arc shots, aerial perspectives, earth zoom-out)
• Lighting Enhancement: Volumetric lighting effects for cinematic quality
• Quality Improvement: Realism boost and upscaling for enhanced video fidelity
• Character Animation: Facial naturalizer and action-specific LoRAs (wink animation)
• FP16 Precision: All models use float16 precision for efficient inference with minimal quality loss

Repository Contents

wan22-fp16-i2v-loras/
└── loras/
    └── wan/
        ├── wan22-action-wink-i2v-v1-low.safetensors        (147 MB)
        ├── wan22-camera-adr1a-v1.safetensors               (293 MB)
        ├── wan22-camera-arcshot-rank16-v2-high.safetensors (293 MB)
        ├── wan22-camera-drone-rank16-v2.safetensors        (293 MB)
        ├── wan22-camera-earthzoomout.safetensors           (293 MB)
        ├── wan22-camera-rotation-rank16-v2.safetensors     (293 MB)
        ├── wan22-face-naturalizer.safetensors              (586 MB)
        ├── wan22-light-volumetric.safetensors              (293 MB)
        └── wan22-upscale-realismboost-t2v-14b.safetensors  (293 MB)

Total Repository Size: 2.8 GB

Model Files

Hardware Requirements

Minimum Requirements

• VRAM: 12 GB (for single LoRA usage with WAN base model)
• RAM: 16 GB system memory
• Disk Space: 3 GB for LoRA collection
• GPU: NVIDIA RTX 3060 (12GB) or equivalent

Recommended Requirements

• VRAM: 24 GB (for multiple simultaneous LoRAs)
• RAM: 32 GB system memory
• Disk Space: 5 GB (with base model cache)
• GPU: NVIDIA RTX 4090, A6000, or equivalent

Base Model Requirements

These LoRAs require the WAN v2.2 base model (separate download):

• WAN base model: ~14 GB additional disk space
• Combined VRAM usage: 16-24 GB depending on configuration

Usage Examples

Basic LoRA Loading with Diffusers

from diffusers import WANPipeline
import torch

# Load base WAN pipeline
pipe = WANPipeline.from_pretrained(
    "Wanniwatch/WAN22",
    torch_dtype=torch.float16,
    variant="fp16"
).to("cuda")

# Load camera rotation LoRA
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-camera-rotation-rank16-v2.safetensors",
    adapter_name="camera_rotation"
)

# Generate video with camera rotation
prompt = "A majestic lion on a savanna, cinematic camera rotation around subject"
video = pipe(
    prompt=prompt,
    num_frames=64,
    height=512,
    width=512,
    guidance_scale=7.5,
    lora_scale=0.8  # Adjust strength (0.0-1.0)
).frames[0]

Combining Multiple LoRAs

# Load multiple LoRAs for enhanced control
lora_configs = [
    ("wan22-camera-drone-rank16-v2.safetensors", "drone", 0.7),
    ("wan22-light-volumetric.safetensors", "lighting", 0.6),
    ("wan22-upscale-realismboost-t2v-14b.safetensors", "quality", 0.5)
]

for weight_name, adapter_name, scale in lora_configs:
    pipe.load_lora_weights(
        "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
        weight_name=weight_name,
        adapter_name=adapter_name
    )

# Set adapter weights
pipe.set_adapters(
    ["drone", "lighting", "quality"],
    adapter_weights=[0.7, 0.6, 0.5]
)

prompt = "Aerial drone shot descending over a misty forest at dawn, volumetric god rays"
video = pipe(prompt=prompt, num_frames=64, height=768, width=768).frames[0]

Image-to-Video with Action LoRA

from PIL import Image

# Load image-to-video pipeline
pipe = WANPipeline.from_pretrained(
    "Wanniwatch/WAN22",
    torch_dtype=torch.float16
).to("cuda")

# Load wink action LoRA
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-action-wink-i2v-v1-low.safetensors",
    adapter_name="wink"
)

# Load input image
image = Image.open("portrait.jpg")

# Generate video with wink animation
video = pipe(
    prompt="Portrait, person winking at camera",
    image=image,
    num_frames=32,
    lora_scale=0.6
).frames[0]

Advanced Camera Control

# Earth zoom-out effect
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-camera-earthzoomout.safetensors"
)

prompt = "Starting from street level, camera zooms out revealing city, country, Earth from space"
video = pipe(
    prompt=prompt,
    num_frames=96,  # Longer sequence for zoom effect
    height=1024,
    width=1024,
    lora_scale=0.85
).frames[0]

Cinematic Lighting Enhancement

# Volumetric lighting
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-light-volumetric.safetensors"
)

prompt = "Dark cathedral interior, volumetric light beams through stained glass windows"
video = pipe(
    prompt=prompt,
    num_frames=64,
    guidance_scale=8.5,  # Higher guidance for dramatic lighting
    lora_scale=0.9
).frames[0]

Quality Enhancement Pipeline

# Realism boost for text-to-video
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-upscale-realismboost-t2v-14b.safetensors"
)

# Combine with face naturalizer for character videos
pipe.load_lora_weights(
    "E:/huggingface/wan22-fp16-i2v-loras/loras/wan",
    weight_name="wan22-face-naturalizer.safetensors",
    adapter_name="face"
)

prompt = "Close-up of a person smiling naturally in golden hour lighting"
video = pipe(
    prompt=prompt,
    num_frames=48,
    height=768,
    width=512,
    lora_scale=0.7
).frames[0]

Model Specifications

Architecture

• Base Model: WAN v2.2 (Wanniwatch video generation model)
• LoRA Rank: Rank 16 for most camera/lighting LoRAs
• Adapter Type: Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning
• Precision: FP16 (float16) for all models
• Format: SafeTensors (secure tensor serialization)

LoRA Categories

Camera Control (5 LoRAs)

• Rotation: 360° camera rotation around subjects
• Drone: Aerial drone cinematography movements
• Arc Shot: Curved/circular camera paths around subjects
• ADR1A: Advanced dynamic camera control system
• Earth Zoom: Extreme zoom-out from ground to orbital perspective

Quality Enhancement (3 LoRAs)

• Realism Boost: Enhanced photorealism for text-to-video generation
• Face Naturalizer: Improved facial animation quality and natural expressions
• Upscale: Quality enhancement for higher resolution output

Effects (1 LoRA)

• Volumetric Lighting: Atmospheric lighting with light shafts and fog effects

Actions (1 LoRA)

• Wink Animation: Character winking for image-to-video generation

Technical Details

• All LoRAs use rank-16 decomposition (except action LoRA which is lower rank)
• Compatible with diffusers library version ≥0.25.0
• Supports both text-to-video and image-to-video pipelines
• LoRA scaling factor: 0.5-1.0 (adjustable per use case)

Performance Tips and Optimization

LoRA Scaling Guidelines

• Camera LoRAs: 0.7-0.9 for pronounced effects, 0.4-0.6 for subtle movements
• Lighting LoRAs: 0.6-0.9 for dramatic effects, 0.3-0.5 for natural lighting
• Quality LoRAs: 0.5-0.8 for balanced enhancement
• Action LoRAs: 0.5-0.7 for controlled character animation

Combining LoRAs

• Maximum Recommended: 3-4 simultaneous LoRAs to avoid conflicts
• Complementary Combinations:
  • Camera + Lighting + Quality (cinematic results)
  • Camera + Face Naturalizer (character focus)
  • Multiple camera LoRAs (complex camera movements)
• Avoid Conflicts: Don't combine multiple quality enhancement LoRAs simultaneously

Memory Optimization

• Use torch.float16 for all operations
• Enable xformers memory efficient attention: pipe.enable_xformers_memory_efficient_attention()
• Reduce resolution for testing: Start with 512x512 before scaling to 1024x1024
• Process fewer frames: 32-64 frames for testing, 96+ for final renders
• Unload unused LoRAs: pipe.unload_lora_weights() between generations

Inference Speed

• Expected generation time (RTX 4090, 64 frames, 768x768): 45-60 seconds
• Single LoRA overhead: ~5-10% slower than base model
• Multiple LoRAs: ~10-20% slower than base model
• Batch processing: Not recommended due to VRAM constraints

Prompt Engineering

• Camera LoRAs: Include camera movement keywords (e.g., "rotating camera", "drone shot")
• Lighting LoRAs: Specify lighting conditions (e.g., "volumetric rays", "god beams")
• Quality LoRAs: Focus on detail keywords (e.g., "photorealistic", "highly detailed")
• Action LoRAs: Explicitly describe the action (e.g., "person winking")

License

This model collection is released under the WAN License (other). Please refer to the official WAN v2.2 license terms from Wanniwatch for usage restrictions and commercial licensing.

License Terms Summary

• Research and non-commercial use permitted
• Commercial use may require separate licensing
• Redistribution must preserve original attribution
• No warranty or liability provided

For complete license details, visit: Wanniwatch/WAN22 on Hugging Face

Citation

If you use these LoRAs in your research or projects, please cite:

@misc{wan22-loras,
  title={WAN v2.2 LoRA Collection for Video Generation},
  author={Wanniwatch Team},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Wanniwatch/WAN22}}
}

Related Resources

Official Links

• Base Model: Wanniwatch/WAN22
• Documentation: WAN v2.2 Official Docs
• Community: WAN Discord Server

Recommended Complementary Models

• VAE: Enhanced video VAE for improved quality
• Upscalers: RealESRGAN for post-processing enhancement
• Controlnets: Depth/pose control for structured video generation

Additional LoRA Collections

• WAN v2.2 FP8 LoRAs (lower precision, faster inference)
• WAN v2.2 Style LoRAs (artistic styles and aesthetics)
• WAN v2.2 Motion LoRAs (specialized motion patterns)

Troubleshooting

Common Issues

Out of Memory Errors:

• Reduce resolution or frame count
• Use single LoRA instead of multiple
• Enable memory efficient attention
• Close other GPU applications

Low Quality Output:

• Increase guidance scale (7.5-9.0)
• Adjust LoRA scale (try 0.7-0.9)
• Use quality enhancement LoRA
• Increase resolution if VRAM allows

LoRA Not Loading:

• Verify absolute file path is correct
• Check diffusers version (≥0.25.0)
• Ensure base model is loaded correctly
• Confirm safetensors format compatibility

Unexpected Camera Movements:

• Lower LoRA scale for subtler effects
• Refine prompt with specific camera keywords
• Avoid conflicting camera LoRAs simultaneously

Contact and Support

For questions, issues, or contributions:

• Issues: Report technical problems via GitHub Issues
• Community: Join the WAN Discord for community support
• Email: [email protected] for commercial inquiries

---

Last Updated: October 2025 README Version: v1.4 Model Version: WAN v2.2 LoRA Collection: 9 specialized adapters