wan25-fp8-i2v

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

WAN 2.5 FP8 Image-to-Video (I2V) - Video Generation Model (Placeholder Repository)

This directory is reserved for WAN 2.5 FP8 Image-to-Video (I2V) models and components. Currently a placeholder repository prepared for future model storage.

Model Description

WAN (World Animation Network) 2.5 is an anticipated next-generation video generation model. This FP8 I2V variant focuses specifically on animating static images into high-quality video sequences with reduced memory requirements while maintaining excellent temporal coherence and motion quality.

Current Status: 🟡 Placeholder Repository - No model files present

Expected Capabilities (When Released)

• Image-to-Video (I2V): Primary focus - animate static images into video sequences
• Motion Control: Fine-grained control over motion intensity and direction
• Camera Control: Advanced camera motion and positioning controls
• Temporal Coherence: Smooth frame-to-frame consistency for realistic animation
• High Resolution: Support for high-resolution video generation (up to 1280×720)
• FP8 Precision: Memory-efficient 8-bit floating point format for broader accessibility

Repository Contents

wan25-fp8-i2v/
├── diffusion_models/
│   └── wan/              # Base WAN 2.5 I2V models (empty)
└── README.md             # This file (15KB)

Current Repository Size: ~15KB (documentation only)

Expected Structure (When Populated)

wan25-fp8-i2v/
├── diffusion_models/
│   └── wan/
│       ├── wan25-i2v-14b-fp8-high-scaled.safetensors  # I2V model (high quality) (~16-20GB)
│       └── wan25-i2v-14b-fp8-low-scaled.safetensors   # I2V model (low quality) (~16-20GB)
└── config/
    ├── model_config.json                              # Model configuration (<1MB)
    └── pipeline_config.json                           # Pipeline settings (<1MB)

Expected Total Size: 32-40GB for complete I2V model collection

Related Repositories:

• wan25-fp8-i2v-loras/: Camera control and enhancement LoRAs for WAN 2.5 I2V
• wan25-vae/: Shared VAE encoder/decoder for WAN 2.5 models

Hardware Requirements

Estimated Requirements (When Available)

FP8 Precision I2V Models:

• VRAM: 10-12GB minimum (low-scaled), 16GB recommended (high-scaled), 24GB for optimal performance
• Disk Space: 32-40GB for complete I2V model collection (both quality variants)
• System RAM: 32GB recommended, 16GB minimum
• GPU: NVIDIA RTX 4070 Ti or higher (RTX 4090, RTX 6000 Ada recommended) with FP8 support
• CUDA: 12.0+ for optimal FP8 performance
• Operating System: Windows 10/11, Linux (Ubuntu 20.04+)

Benefits of FP8:

• ~50% VRAM reduction vs FP16
• Faster inference on Ada Lovelace/Hopper architectures
• Minimal quality loss with proper calibration
• Enables larger batch sizes or higher resolutions

Usage Examples

Basic Image-to-Video Generation (Speculative)

from diffusers import DiffusionPipeline, AutoencoderKL
from PIL import Image
import torch

# Load WAN 2.5 FP8 I2V model (when released)
pipe = DiffusionPipeline.from_single_file(
    "E:/huggingface/wan25-fp8-i2v/diffusion_models/wan/wan25-i2v-14b-fp8-high-scaled.safetensors",
    torch_dtype=torch.float8_e4m3fn,  # FP8 precision
)

# Load shared VAE from wan25-vae repository
pipe.vae = AutoencoderKL.from_single_file(
    "E:/huggingface/wan25-vae/vae/wan25-vae-fp8.safetensors",
    torch_dtype=torch.float8_e4m3fn
)

pipe.to("cuda")

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

# Generate video from static image
video = pipe(
    image=input_image,
    prompt="The scene comes to life with gentle, natural movement",
    num_frames=48,              # 2 seconds at 24fps
    num_inference_steps=40,
    guidance_scale=6.5,
    motion_intensity=0.7        # Control motion strength (0-1)
).frames

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

With Camera Control LoRA (Speculative)

# Load camera control LoRA from separate repository
pipe.load_lora_weights(
    "E:/huggingface/wan25-fp8-i2v-loras/loras/camera_control_v3.safetensors"
)

# Load landscape image
landscape_image = Image.open("mountain_scene.jpg")

# Animate with camera motion
video = pipe(
    image=landscape_image,
    prompt="Cinematic pan across the mountain vista",
    num_frames=48,
    camera_motion="pan_right",      # Camera pan right
    camera_speed=0.5,               # Moderate speed
    motion_intensity=0.6
).frames

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

Quality-Scaled Model Selection (Speculative)

# Use low-scaled variant for faster inference (lower VRAM)
pipe_low = DiffusionPipeline.from_single_file(
    "E:/huggingface/wan25-fp8-i2v/diffusion_models/wan/wan25-i2v-14b-fp8-low-scaled.safetensors",
    torch_dtype=torch.float8_e4m3fn
)

# Use high-scaled variant for maximum quality
pipe_high = DiffusionPipeline.from_single_file(
    "E:/huggingface/wan25-fp8-i2v/diffusion_models/wan/wan25-i2v-14b-fp8-high-scaled.safetensors",
    torch_dtype=torch.float8_e4m3fn
)

# Generate with high quality model
portrait_image = Image.open("portrait.jpg")
video = pipe_high(
    image=portrait_image,
    prompt="Subtle facial expressions and natural head movement",
    num_frames=48,
    guidance_scale=7.0
).frames

Model Specifications

Architecture (Expected)

• Base Architecture: Latent diffusion with 3D temporal attention (I2V-specific)
• Model Size: 14 billion parameters
• Precision: FP8 E4M3 (8-bit floating point)
• Format: SafeTensors (secure, efficient)
• Input: Static images + text prompts (guidance)
• Output: Video sequences (animated from input image)
• Frame Rate: 24 FPS (default)
• Resolution: Up to 1280×720 (720p), input image dependent
• Temporal Consistency: Advanced temporal attention with I2V conditioning
• Quality Variants: High-scaled and low-scaled models for quality/VRAM trade-offs

Performance Optimization

FP8 Advantages:

• Reduced memory footprint for longer videos
• Faster inference on modern GPUs (Ada/Hopper)
• Near-FP16 quality with proper quantization
• Better multi-GPU scaling
• Energy efficiency improvements

Recommended Settings:

• Use FP8 E4M3 format for maximum compatibility
• Enable flash attention for faster inference
• Use gradient checkpointing for memory efficiency
• Optimize batch size based on VRAM availability

Performance Tips

When Models Become Available

1. GPU Selection: Use NVIDIA RTX 4090 or newer for native FP8 support
2. Memory Management: Enable enable_model_cpu_offload() for VRAM constrained systems
3. Batch Processing: Process multiple prompts in batches when VRAM allows
4. VAE Tiling: Use enable_vae_tiling() for high-resolution generation
5. Flash Attention: Enable for 2-3x speedup: pipe.enable_xformers_memory_efficient_attention()
6. Compilation: Use torch.compile() for additional 10-20% speedup

Quality Optimization

• Inference Steps: 30-50 steps for balanced quality/speed
• Guidance Scale: 7.0-9.0 for text-to-video, 5.0-7.0 for image-to-video
• Resolution: Start with 512×512, upscale to target resolution
• Frame Count: Generate shorter clips (24-48 frames), stitch if needed
• Prompt Engineering: Use detailed, specific prompts for best results

Related WAN Models

While WAN 2.5 FP8 I2V is not yet available, explore existing WAN I2V models:

Available I2V Alternatives

• WAN 2.1 I2V (480p/720p): FP16/FP8 image-to-video models with v1 camera control LoRAs
• WAN 2.2 I2V: FP16/FP8 image-to-video models with enhanced v2 camera controls
• WAN LightX2V I2V: Distilled I2V models (14B parameters) for faster inference
• WAN 2.1/2.2 VAE: Shared video VAE encoder/decoder

I2V Version Comparison

License

WAN models typically use a custom license. Please review the specific license terms when models are released.

Expected License Restrictions:

• Usage terms (commercial/non-commercial) to be determined
• Attribution requirements
• Redistribution limitations
• Content generation guidelines
• Regional restrictions (if any)

Important: Always review the official license file when downloading WAN models.

Citation

When WAN 2.5 I2V models become available, cite as:

@misc{wan25-i2v-fp8,
  title={WAN 2.5 I2V: Advanced Image-to-Video Generation with FP8 Optimization},
  author={WAN Team},
  year={2025},
  howpublished={\url{https://huggingface.co/wan-models/wan-2.5-fp8-i2v}},
  note={FP8-optimized image-to-video generation model with 14B parameters}
}

Links and Resources

Official Resources (when available):

• Model Hub: https://huggingface.co/wan-models/
• Documentation: Check official WAN documentation
• Community: Hugging Face forums and Discord
• Paper: To be published

Related Technologies:

• Diffusers Library
• SafeTensors Format
• FP8 Training Guide
• Hugging Face Model Hub

Stay Updated

Monitoring WAN 2.5 Release:

1. Watch official WAN model repositories on Hugging Face
2. Follow model authors and organizations
3. Subscribe to community announcement channels
4. Check this repository for updates

How to Prepare:

• Ensure GPU drivers support FP8 operations (CUDA 12.0+)
• Install latest diffusers library: pip install -U diffusers
• Review existing WAN 2.2 implementations for compatibility
• Test FP8 inference with available models
• Prepare adequate disk space (80GB recommended)

---

Changelog

v1.4 (Current)

• Validation Update: Verified complete compliance with HuggingFace metadata standards
• YAML Frontmatter: Confirmed proper structure with all required fields (license, library_name, pipeline_tag, tags)
• Quality Assurance: Validated against comprehensive README quality checklist
• Documentation Completeness: Ensured all sections meet production-ready standards

v1.3

• I2V-Focused Update: Specialized documentation for Image-to-Video variant
• Repository Structure: Updated to reflect I2V-specific file organization (high/low-scaled models)
• Usage Examples: Completely revised with I2V-first examples and quality-scaled model selection
• Hardware Requirements: Adjusted for I2V-specific VRAM and disk space needs
• Model Specifications: Added 14B parameter details and I2V-specific architecture notes
• Version Comparison: Updated table to focus on I2V model evolution across WAN versions
• Related Repositories: Documented integration with wan25-vae and wan25-fp8-i2v-loras

v1.2

• Updated to README Version v1.2 format
• Enhanced usage examples with image-to-video
• Improved hardware requirements section
• Added more comprehensive model specifications
• Expanded performance tips and quality optimization

v1.1

• Updated YAML frontmatter to match HuggingFace standards
• Simplified tags for better discoverability
• Moved version header after YAML block
• Maintained comprehensive documentation

v1.0

• Initial placeholder repository structure
• Comprehensive documentation prepared
• Hugging Face metadata configured
• Usage examples and specifications documented

---

Repository Status: 🟡 Placeholder - Awaiting WAN 2.5 FP8 I2V model release

Last Updated: 2025-10-28

Note: This is currently an empty placeholder directory with documentation only. Model files will be added when WAN 2.5 FP8 Image-to-Video models are officially released. This repository is specifically for I2V models; for Text-to-Video variants, see separate wan25-fp8-t2v repositories. Check back regularly for updates or watch this repository for changes.