wangkanai / wan25-fp16-loras

WAN 2.5 FP16 LoRA Collection

High-quality LoRA adapters for WAN 2.5 video generation models in FP16 precision for enhanced control and quality improvements.

Model Description

This repository contains LoRA (Low-Rank Adaptation) adapters specifically designed for the WAN 2.5 video generation model. These adapters enable fine-grained control over various aspects of video generation including camera movements, lighting conditions, and overall quality enhancements without requiring full model retraining.

Key Features :

• Camera Control LoRAs : Precise control over camera movements (pan, tilt, zoom, dolly)
• Lighting Enhancement : Dynamic lighting adjustment and atmospheric control
• Quality Improvements : Enhanced detail, reduced artifacts, improved temporal consistency
• FP16 Precision : Balance between quality and performance
• Modular Design : Mix and match LoRAs for combined effects

Use Cases :

• Cinematic video generation with professional camera movements
• Lighting-controlled scene generation
• Quality enhancement for existing generations
• Style transfer and artistic effects
• Motion control and temporal consistency improvements

Repository Contents

Current Status : Repository structure is initialized and ready for model files. No LoRA files are currently present in the repository.

Directory Structure :

wan25-fp16-loras/
├── README.md                           # This file
└── loras/
    └── wan/
        ├── camera/                     # Camera control LoRAs
        │   ├── pan_left_right.safetensors
        │   ├── tilt_up_down.safetensors
        │   ├── zoom_in_out.safetensors
        │   └── dolly_forward_back.safetensors
        ├── lighting/                   # Lighting enhancement LoRAs
        │   ├── natural_light.safetensors
        │   ├── dramatic_lighting.safetensors
        │   └── ambient_occlusion.safetensors
        └── quality/                    # Quality enhancement LoRAs
            ├── detail_enhancement.safetensors
            ├── temporal_consistency.safetensors
            └── artifact_reduction.safetensors

Current Repository Size : 18 KB (structure only)

Expected File Sizes (when populated):

• Camera control LoRAs: ~50-150 MB each
• Lighting LoRAs: ~75-200 MB each
• Quality enhancement LoRAs: ~100-250 MB each
• Total Repository Size : ~1-2 GB (when fully populated with all LoRAs)

Hardware Requirements

Minimum Requirements

• VRAM : 16 GB (for inference with single LoRA)
• RAM : 16 GB system memory
• Disk Space : 5 GB (includes base model + LoRAs)
• GPU : NVIDIA RTX 3060 12GB or equivalent

Recommended Requirements

• VRAM : 24 GB (for multiple LoRAs and longer sequences)
• RAM : 32 GB system memory
• Disk Space : 10 GB (for full collection + workspace)
• GPU : NVIDIA RTX 4090, A6000, or equivalent

Optimal Performance

• VRAM : 48 GB+ (A6000, A100)
• RAM : 64 GB system memory
• Disk Space : 20 GB (includes variations and checkpoints)
• GPU : NVIDIA A100 80GB or H100

Usage Examples

Loading Base Model with LoRA

import torch
from diffusers import WanPipeline

# Load base WAN 2.5 model
pipe = WanPipeline.from_pretrained(
    "E:/huggingface/wan25-fp16",  # Base model path
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe.to("cuda")

# Load camera control LoRA
pipe.load_lora_weights(
    "E:/huggingface/wan25-fp16-loras/loras/wan/camera",
    weight_name="pan_left_right.safetensors",
    adapter_name="camera_pan"
)

# Generate video with camera pan effect
prompt = "A cinematic landscape scene with smooth camera pan"
video = pipe(
    prompt=prompt,
    num_frames=48,
    height=512,
    width=768,
    num_inference_steps=30,
    guidance_scale=7.5,
    cross_attention_kwargs={"scale": 0.8}  # LoRA strength
).frames[0]

# Save video
from diffusers.utils import export_to_video
export_to_video(video, "output_with_camera_pan.mp4", fps=24)

Combining Multiple LoRAs

import torch
from diffusers import WanPipeline

# Load base model
pipe = WanPipeline.from_pretrained(
    "E:/huggingface/wan25-fp16",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe.to("cuda")

# Load multiple LoRAs
lora_configs = [
    {
        "path": "E:/huggingface/wan25-fp16-loras/loras/wan/camera",
        "weight_name": "dolly_forward_back.safetensors",
        "adapter_name": "camera_dolly",
        "scale": 0.7
    },
    {
        "path": "E:/huggingface/wan25-fp16-loras/loras/wan/lighting",
        "weight_name": "dramatic_lighting.safetensors",
        "adapter_name": "lighting_dramatic",
        "scale": 0.6
    },
    {
        "path": "E:/huggingface/wan25-fp16-loras/loras/wan/quality",
        "weight_name": "detail_enhancement.safetensors",
        "adapter_name": "quality_detail",
        "scale": 0.5
    }
]

# Load all LoRAs
for config in lora_configs:
    pipe.load_lora_weights(
        config["path"],
        weight_name=config["weight_name"],
        adapter_name=config["adapter_name"]
    )

# Set adapter scales
adapter_names = [cfg["adapter_name"] for cfg in lora_configs]
adapter_scales = [cfg["scale"] for cfg in lora_configs]
pipe.set_adapters(adapter_names, adapter_scales)

# Generate with combined effects
prompt = "A dramatic scene with forward camera movement and enhanced details"
video = pipe(
    prompt=prompt,
    num_frames=96,
    height=576,
    width=1024,
    num_inference_steps=40,
    guidance_scale=8.0
).frames[0]

export_to_video(video, "output_combined_loras.mp4", fps=24)

Dynamic LoRA Strength Control

import torch
from diffusers import WanPipeline

pipe = WanPipeline.from_pretrained(
    "E:/huggingface/wan25-fp16",
    torch_dtype=torch.float16
)
pipe.to("cuda")

# Load quality enhancement LoRA
pipe.load_lora_weights(
    "E:/huggingface/wan25-fp16-loras/loras/wan/quality",
    weight_name="temporal_consistency.safetensors",
    adapter_name="temporal"
)

# Test different LoRA strengths
strengths = [0.3, 0.5, 0.7, 0.9]
prompt = "Smooth flowing water in a mountain stream"

for strength in strengths:
    video = pipe(
        prompt=prompt,
        num_frames=48,
        height=512,
        width=768,
        num_inference_steps=30,
        guidance_scale=7.5,
        cross_attention_kwargs={"scale": strength}
    ).frames[0]

    export_to_video(video, f"output_strength_{strength}.mp4", fps=24)
    print(f"Generated video with LoRA strength: {strength}")

Unloading LoRAs

# Unload specific LoRA
pipe.unload_lora_weights()

# Or disable specific adapter
pipe.disable_lora()

# To enable again
pipe.enable_lora()

Model Specifications

LoRA Architecture

• Format : SafeTensors (.safetensors)
• Precision : FP16 (16-bit floating point)
• Rank : Typically 32-64 (configurable)
• Target Modules : Attention layers, cross-attention, temporal layers
• Compatibility : WAN 2.5 base model (FP16 and FP8 variants)

LoRA Categories

Camera Control :

• Pan (left/right horizontal movement)
• Tilt (up/down vertical movement)
• Zoom (in/out focal length)
• Dolly (forward/backward position)
• Roll (rotation around lens axis)

Lighting Enhancement :

• Natural lighting (sun, sky, ambient)
• Dramatic lighting (high contrast, spotlights)
• Ambient occlusion (shadows and depth)
• Color temperature control
• HDR enhancement

Quality Improvements :

• Detail enhancement (texture, sharpness)
• Temporal consistency (frame coherence)
• Artifact reduction (blocking, flickering)
• Motion blur control
• Noise reduction

Technical Details

• Base Model : WAN 2.5 (black-forest-labs/wan-2.5)
• Training Data : Curated video datasets with specific camera/lighting/quality attributes
• Training Framework : Diffusers with LoRA training scripts
• Optimization : Flash Attention 2, gradient checkpointing
• Validation : Temporal consistency metrics, perceptual quality scores

Performance Tips and Optimization

LoRA Strength Guidelines

• Camera Control : 0.6-0.9 for strong effects, 0.3-0.5 for subtle movements
• Lighting : 0.5-0.7 for natural adjustments, 0.7-0.9 for dramatic effects
• Quality : 0.4-0.6 for enhancement without over-processing

Memory Optimization

# Enable memory-efficient attention
pipe.enable_xformers_memory_efficient_attention()

# Use CPU offloading for limited VRAM
pipe.enable_sequential_cpu_offload()

# Reduce precision for inference (if needed)
pipe.vae.to(dtype=torch.float16)
pipe.unet.to(dtype=torch.float16)

Batch Processing

# Process multiple prompts efficiently
prompts = [
    "Scene 1 with camera pan",
    "Scene 2 with dramatic lighting",
    "Scene 3 with enhanced details"
]

videos = pipe(
    prompt=prompts,
    num_frames=48,
    height=512,
    width=768,
    num_inference_steps=30,
    guidance_scale=7.5
).frames

for i, video in enumerate(videos):
    export_to_video(video, f"batch_output_{i}.mp4", fps=24)

Quality vs Speed Trade-offs

• Fast : 20-25 steps, lower resolution (512x512), single LoRA
• Balanced : 30-35 steps, medium resolution (768x512), 1-2 LoRAs
• High Quality : 40-50 steps, high resolution (1024x576), multiple LoRAs

Recommended Combinations

1. Cinematic : Camera dolly + Dramatic lighting + Detail enhancement
2. Natural : Camera pan + Natural lighting + Temporal consistency
3. Artistic : Camera zoom + Color temperature + Quality enhancement

License

This repository contains LoRA adapters for the WAN 2.5 model. Please refer to the original WAN model license for terms and conditions.

License Type : Custom license (see WAN model documentation)

Usage Restrictions :

• Review base model license terms before commercial use
• LoRA weights may have additional restrictions
• Respect content policy and ethical guidelines

Attribution : When using these LoRAs in published work, please cite both the base WAN model and this LoRA collection.

Citation

If you use these LoRA adapters in your research or applications, please cite:

@misc{wan25-fp16-loras,
  title={WAN 2.5 FP16 LoRA Collection},
  author={[To be determined based on actual model creators]},
  year={2025},
  howpublished={\url{https://huggingface.co/[your-username]/wan25-fp16-loras}},
  note={LoRA adapters for WAN 2.5 video generation model}
}

@misc{wan25,
  title={WAN 2.5: Advanced Video Generation Model},
  author={Black Forest Labs},
  year={2025},
  howpublished={\url{https://blackforestlabs.ai/}}
}

Resources and Support

Official Resources

• WAN Model Documentation : https://blackforestlabs.ai/wan
• Diffusers Documentation : https://huggingface.co/docs/diffusers
• LoRA Training Guide : https://huggingface.co/docs/diffusers/training/lora

Community and Support

• Hugging Face Hub : https://huggingface.co/models?library=diffusers&pipeline_tag=text-to-video
• Diffusers GitHub : https://github.com/huggingface/diffusers
• Community Forums : https://discuss.huggingface.co/

Related Models

• WAN 2.5 Base Model : E:/huggingface/wan25-fp16
• WAN VAE : E:/huggingface/wan25-fp16/vae
• FLUX Models : E:/huggingface/flux-dev-fp16

Version History

v1.1 (2025-10-14)

• Updated README with accurate repository status
• Clarified that no LoRA files are currently present
• Added current repository size information
• Enhanced documentation clarity

v1.0 (2025-10-13)

• Initial repository structure
• Comprehensive documentation and usage examples
• Placeholder for camera, lighting, and quality LoRAs
• Ready for model file integration

Contributing

To add new LoRAs to this collection:

1. Organize by category : Place in camera/, lighting/, or quality/ subdirectories
2. Use SafeTensors format : Convert to .safetensors for security and efficiency
3. Update README : Document new LoRAs with descriptions and usage examples
4. Test compatibility : Verify with WAN 2.5 base model
5. Provide examples : Include sample code and recommended settings

Important Notes

Repository Status :

• Directory structure is initialized and ready for model files
• No LoRA model files are currently present (0 files)
• Examples and usage documentation are provided for future use
• Structure follows WAN LoRA organization conventions

Usage Information :

• All code examples use absolute Windows paths (E:/huggingface/...)
• Adjust paths according to your local setup if different
• LoRA adapters are modular and can be used independently or combined
• Experiment with different strength values (0.3-0.9) to achieve desired effects
• Model files will be added as they become available from official sources