wan21-lightx2v-i2v-14b-720p

<!-- README Version: v1.5 -->

WAN LightX2V T2V LoRA Adapters (720p)

High-quality LoRA (Low-Rank Adaptation) adapters for the LightX2V 14B text-to-video generation model at 720p resolution. These adapters enable efficient fine-tuning and accelerated inference through CFG (Classifier-Free Guidance) step distillation.

Model Description

This repository contains 5 CFG step-distilled LoRA adapters designed to accelerate text-to-video generation while maintaining high quality output at 720p resolution. The adapters are available in multiple ranks (8, 16, 32, 64, 128) to accommodate different hardware configurations and quality requirements.

Key Features

• Multiple Rank Options: Choose from 5 different ranks (8-128) for flexibility
• CFG Step Distillation: Reduces inference steps from 50-100 down to 15-30 steps
• BF16 Precision: Brain floating point format for stability and efficiency
• 720p Optimized: Designed for 1280x720 resolution video generation
• Fast Inference: 2-3x speedup compared to non-distilled models
• SafeTensors Format: Secure and efficient model format

Repository Contents

wan21-lightx2v-i2v-14b-720p/
└── loras/
    └── wan/
        ├── wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank8-bf16.safetensors (82MB)
        ├── wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank16-bf16.safetensors (156MB)
        ├── wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank32-bf16.safetensors (305MB)
        ├── wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank64-bf16.safetensors (602MB)
        └── wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank128-bf16.safetensors (1.2GB)

Total Repository Size: ~2.3GB (all 5 adapters combined)

File Details

Hardware Requirements

Minimum Configuration (Rank 8-16)

• GPU: NVIDIA RTX 3060 (12GB VRAM) or equivalent
• System RAM: 16GB
• Storage: 500MB for adapters + base model storage
• GPU Architecture: Ampere or newer (for BF16 support)
• CUDA: 11.8+ or 12.1+

Recommended Configuration (Rank 32-64)

• GPU: NVIDIA RTX 4070 Ti (16GB VRAM) or RTX 3090 (24GB)
• System RAM: 32GB
• Storage: 1GB for adapters + base model storage
• Resolution: Optimized for 720p (1280x720)
• OS: Windows 10/11, Linux (Ubuntu 20.04+)

High-End Configuration (Rank 128)

• GPU: NVIDIA RTX 4090 (24GB VRAM) or A100 (40GB)
• System RAM: 64GB
• Storage: 1.5GB for adapters + base model storage
• Use Case: Maximum quality production workflows

VRAM Usage Estimates (720p, 24 frames)

Disk Space Requirements

• Individual Adapter: 82MB - 1.2GB (depending on rank)
• All Adapters: ~2.3GB total
• Base Model: ~28GB (LightX2V T2V 14B - not included)
• Total Space Needed: ~30GB (base model + all adapters + workspace)

Usage Examples

Basic Text-to-Video with Diffusers

from diffusers import DiffusionPipeline
import torch

# Load base T2V model
pipe = DiffusionPipeline.from_pretrained(
    "lightx2v/lightx2v-t2v-14b",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter (rank-32 recommended for balanced quality/performance)
pipe.load_lora_weights(
    "E:/huggingface/wan21-lightx2v-i2v-14b-720p/loras/wan/wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank32-bf16.safetensors"
)

# Generate video from text prompt
prompt = "A serene mountain landscape at sunset with golden light, cinematic camera movement, 720p quality"
video = pipe(
    prompt=prompt,
    num_inference_steps=20,  # Reduced steps thanks to CFG distillation
    guidance_scale=7.5,
    num_frames=24,
    height=720,
    width=1280
).frames

# Save video to file
from diffusers.utils import export_to_video
export_to_video(video, "output_t2v_720p.mp4", fps=8)

Rank Selection Based on Hardware

import os

# Define base path to LoRA adapters
LORA_PATH = "E:/huggingface/wan21-lightx2v-i2v-14b-720p/loras/wan"

# Select rank based on your hardware:
# - Rank 8-16: Budget GPUs (12GB VRAM)
# - Rank 32: Recommended for most users (16-24GB VRAM)
# - Rank 64-128: High-end GPUs for maximum quality (24GB+ VRAM)
rank = 32

lora_file = f"{LORA_PATH}/wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank{rank}-bf16.safetensors"
pipe.load_lora_weights(lora_file)

# Generate with optimized settings
video = pipe(
    prompt="Aerial drone shot rising above a misty forest at sunrise, cinematic",
    num_inference_steps=20,
    guidance_scale=7.5,
    num_frames=24,
    height=720,
    width=1280
).frames

Advanced: Memory Optimization

from diffusers import DiffusionPipeline
import torch

# Load with memory optimizations
pipe = DiffusionPipeline.from_pretrained(
    "lightx2v/lightx2v-t2v-14b",
    torch_dtype=torch.bfloat16,
    variant="bf16"
)

# Enable CPU offloading for lower VRAM usage
pipe.enable_model_cpu_offload()

# Enable attention slicing to reduce VRAM further
pipe.enable_attention_slicing()

# Load lower rank for constrained environments
pipe.load_lora_weights(
    "E:/huggingface/wan21-lightx2v-i2v-14b-720p/loras/wan/wan21-lightx2v-t2v-14b-cfg-step-distill-v2-rank16-bf16.safetensors"
)

# Generate with memory-efficient settings
video = pipe(
    prompt="City street at night with neon lights reflecting on wet pavement",
    num_inference_steps=15,  # Fewer steps = less memory
    guidance_scale=7.0,
    num_frames=16,  # Reduced frames for lower memory
    height=720,
    width=1280
).frames

ComfyUI Integration

1. Copy LoRA file to your ComfyUI loras directory:

   ComfyUI/models/loras/wan/
   └── wan21-lightx2v-t2v-rank32-bf16.safetensors
   

2. Workflow nodes setup:

   • Add "Load LoRA" node
   • Connect to LightX2V T2V model nodes
   • Set LoRA strength: 0.8-1.0

3. Recommended parameters:

   • Steps: 15-25 (distilled model requires fewer)
   • CFG Scale: 6.0-8.0
   • Resolution: 1280x720 (720p)
   • Frames: 16-32 frames

Model Specifications

Architecture Details

• Type: Low-Rank Adaptation (LoRA) for Diffusion Models
• Base Architecture: LightX2V T2V 14B (14 billion parameters)
• Training Method: CFG Step Distillation v2
• Precision: BF16 (Brain Floating Point 16-bit)
• Format: SafeTensors (.safetensors)
• Optimization: Classifier-Free Guidance distillation

Technical Specifications

LoRA Rank Selection Guide

Recommendation: Start with rank-32 for optimal quality/performance balance. Scale up to 64/128 for maximum quality, or down to 16/8 for faster iteration on constrained hardware.

CFG Step Distillation Benefits

• Reduced Steps: 15-30 steps (vs 50-100 for baseline models)
• Speed Improvement: 2-3x faster generation
• Quality Preservation: Maintains visual quality with fewer steps
• CFG Optimization: Better classifier-free guidance behavior
• Consistency: More stable results across different CFG scales
• Cost Efficiency: Lower compute costs for production use

BF16 Format Advantages

• Numerical Stability: Better than FP16, fewer overflow issues
• Dynamic Range: Wider range prevents numerical errors
• Hardware Support: Optimized for NVIDIA Ampere/Ada/Hopper
• Memory Efficient: Half the size of FP32 with minimal quality loss
• Training Stability: Improved gradient stability during fine-tuning

Performance Tips and Optimization

Generation Speed Optimization

1. Use Lower Ranks: Rank 8-32 for faster iteration
2. Reduce Steps: 15-20 steps sufficient with distillation
3. Enable torch.compile(): On PyTorch 2.0+ for JIT compilation
4. CPU Offloading: Use enable_model_cpu_offload() for memory
5. Attention Slicing: enable_attention_slicing() reduces VRAM peaks

Quality Maximization

1. Higher Ranks: Use rank 64 or 128 for best results
2. Optimal Steps: 20-25 steps for 720p quality
3. CFG Scale: 6.5-8.0 range works best
4. Detailed Prompts: Include camera movement, lighting, "720p quality"
5. Frame Count: 24-32 frames for smooth motion

Memory Management

# Aggressive memory optimization
pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()

# Use lower rank
pipe.load_lora_weights("...rank16-bf16.safetensors")

# Reduce batch size and frames
video = pipe(prompt, num_frames=16, height=720, width=1280)

Performance Benchmarks (RTX 4090, Rank 32, 720p)

Benchmarks on RTX 4090, rank-32, 24 frames, 720p resolution. Actual times vary by prompt complexity and system configuration.

Prompting Best Practices

Text-to-Video Prompt Structure

Effective Prompt Template:

[Subject/Scene] + [Action/Movement] + [Camera Work] + [Lighting/Atmosphere] + [Quality Tags]

Example Prompts for 720p

"A majestic eagle soaring through mountain valleys at golden hour, slow motion flight, camera tracking from side, cinematic lighting, 720p HD quality, professional cinematography"

"City street time-lapse with traffic flowing, neon lights reflecting on wet pavement, camera slowly panning right, night scene, high detail, 720p resolution"

"Underwater coral reef with tropical fish swimming, gentle camera movement through scene, clear blue water, sunlight filtering from above, cinematic 720p quality"

"Drone shot rising above a misty forest at sunrise, rays of light breaking through trees, smooth camera ascent from ground level, atmospheric fog, HD quality"

"Close-up of coffee being poured into white cup, slow motion, steam rising, shallow depth of field, warm morning lighting, cinematic 720p"

Prompt Enhancement Tips

• Camera Movement: "dolly zoom", "pan left", "crane shot", "tracking shot", "aerial view"
• Temporal Dynamics: "slow motion", "time-lapse", "real-time", "smooth transition"
• Lighting: "golden hour", "blue hour", "volumetric lighting", "rim lighting"
• Quality Tags: "720p", "HD quality", "cinematic", "professional", "high detail"
• Atmosphere: "misty", "foggy", "atmospheric", "moody", "vibrant"

Prompt Tips for Best Results

• Be Specific: Detailed scene descriptions produce better results
• Include Motion: Describe movement and camera work explicitly
• Mention Resolution: Add "720p" or "HD quality" to prompts
• Use Cinematic Terms: "cinematic", "professional", "broadcast quality"
• Describe Lighting: Lighting dramatically affects video quality
• Keep It Focused: Avoid overly complex multi-scene descriptions

Troubleshooting

Out of Memory (OOM) Errors

Solution 1: Use Lower Rank

# Switch from rank-64 to rank-32 or rank-16
pipe.load_lora_weights("...rank16-bf16.safetensors")

Solution 2: Enable CPU Offloading

pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()

Solution 3: Reduce Frames/Resolution

# Generate shorter clips
video = pipe(prompt, num_frames=16)  # Instead of 24

# Temporarily test with 480p
video = pipe(prompt, height=480, width=854)

Solution 4: Close Other Applications

• Free up VRAM by closing browsers, IDEs, other GPU applications
• Monitor GPU usage with nvidia-smi

Poor Quality Results

Issue: Blurry or Low-Detail Output

• Increase LoRA rank to 64 or 128
• Use 20-25 inference steps instead of 15
• Add "high detail", "sharp focus", "720p quality" to prompts
• Ensure resolution is set to 1280x720

Issue: Inconsistent Motion

• Adjust CFG scale (try 6.5-8.0 range)
• Use more descriptive motion keywords in prompts
• Increase frame count to 24-32 for smoother motion

Issue: Poor Prompt Adherence

• Increase CFG scale to 7.5-8.5
• Make prompts more specific and detailed
• Use rank-32 or higher for better prompt understanding

Slow Generation Speed

Optimization Steps:

1. Use rank-16 or rank-32 instead of rank-128
2. Reduce inference steps to 15-20
3. Enable PyTorch compilation: pipe.unet = torch.compile(pipe.unet)
4. Use xFormers for memory-efficient attention
5. Consider 480p for faster iteration, then upscale

Installation Issues

Missing Dependencies

# Install PyTorch with CUDA 12.1
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# Install required libraries
pip install diffusers transformers accelerate safetensors xformers

# Verify installation
python -c "import torch; print(torch.cuda.is_available())"

BF16 Not Supported

• Requires NVIDIA Ampere architecture or newer (RTX 30/40 series, A100)
• For older GPUs, convert to FP16 or use FP32 (not recommended)

License Information

These LoRA adapters are designed for use with the LightX2V T2V 14B base model. Please ensure compliance with:

• Base Model License: LightX2V T2V 14B license terms
• Adapter License: Follow base model licensing requirements
• Commercial Use: Verify base model allows commercial usage

Important: Always review and comply with the LightX2V base model license before deployment.

Citation

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

@misc{wan21-lightx2v-lora-720p,
  title={WAN LightX2V T2V LoRA Adapters (720p)},
  author={WAN Team},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/wan21-lightx2v-i2v-14b-720p}}
}

@misc{lightx2v-t2v-14b,
  title={LightX2V T2V 14B: Text-to-Video Diffusion Model},
  author={LightX2V Team},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/lightx2v/lightx2v-t2v-14b}}
}

Related Resources

• Base Model: LightX2V T2V 14B
• WAN 2.1 Models: Image-to-video models with camera control
• WAN 2.2 Models: Enhanced I2V/T2V models with advanced features
• 480p I2V LoRAs: wan21-lightx2v-i2v-14b-480p (image-to-video at 480p)
• Diffusers Documentation: Hugging Face Diffusers
• LoRA Documentation: LoRA: Low-Rank Adaptation

Changelog

v1.5 (October 2025)

• Updated to v1.5 with refined YAML metadata per Hugging Face standards
• Simplified tags to core model capabilities (wan, text-to-video, image-generation)
• Removed redundant tags (lora, diffusion, video-generation) per SuperClaude framework guidelines
• Validated YAML frontmatter: proper format, no base_model fields, minimal essential tags
• Maintained comprehensive documentation structure with all technical specifications

v1.4 (October 2024)

• Updated tags to better reflect content: replaced image-generation with video-generation, added lora and diffusion tags
• Improved metadata accuracy for Hugging Face discoverability
• Version bumped to v1.4 for enhanced metadata compliance

v1.3 (October 2024)

• Version update to v1.3 with metadata validation
• Verified YAML frontmatter compliance with Hugging Face standards
• Confirmed all critical requirements met for repository metadata

v1.2 (October 2024)

• Updated YAML frontmatter to remove base_model and base_model_relation fields
• Simplified tags to core categories for better Hugging Face compatibility
• Version bumped to v1.2 for metadata compliance

v1.1 (October 2024)

• Updated README with comprehensive documentation
• Added detailed hardware requirements and VRAM estimates
• Expanded usage examples with memory optimization
• Added troubleshooting section and prompt engineering guide
• Improved YAML frontmatter formatting for Hugging Face compatibility

v1.0 (October 2024)

• Initial release with 5 LoRA adapters (ranks 8, 16, 32, 64, 128)
• CFG step distillation v2 implementation
• BF16 precision for all adapters
• 720p resolution optimization

---

Last Updated: October 2025 Repository Version: v1.5 Total Size: ~2.3GB (5 adapters: ranks 8, 16, 32, 64, 128) Primary Use Case: Text-to-video generation at 720p resolution with accelerated inference