wan21-fp8-loras-nsfw

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

⚠️ WAN 2.1 Action LoRAs (Adult Content)

CONTENT WARNING: This repository contains LoRA adapters trained on adult/NSFW content for video generation. These models are intended for adult users (18+) only and should be used responsibly in accordance with applicable laws and regulations.

Specialized LoRA (Low-Rank Adaptation) adapters for the WAN 2.1 14B image-to-video generation model, focused on specific action sequences and camera angles.

📦 Model Information

• Base Model: WAN 2.1 I2V 14B
• Type: Action-Specific LoRA Adapters
• Version: v1 and v2 variants
• Precision: BF16 (Brain Floating Point 16)
• Content Type: Adult/NSFW
• Total Models: 8 specialized adapters
• Repository Size: 3.2GB

⚠️ Usage Restrictions

• Age Restriction: 18+ only
• Legal Compliance: Users must comply with local laws regarding adult content
• Ethical Use: Not for non-consensual content generation or deepfakes
• Platform Guidelines: Respect platform policies where content is shared
• Content Moderation: Implement appropriate content warnings and filters

📁 Repository Contents

wan21-fp8-loras-nsfw/
└── loras/
    └── wan/
        ├── wan21-action-bounce-v1.safetensors (343MB)
        ├── wan21-action-doggypov-v2.safetensors (514MB)
        ├── wan21-action-doggystylefromfront-t2v.safetensors (293MB)
        ├── wan21-action-front-doggy-plow-v2.safetensors (514MB)
        ├── wan21-action-missionary-pov-i2v-v2.safetensors (343MB)
        ├── wan21-action-missionary-pov-v2.safetensors (293MB)
        ├── wan21-action-orgasm-i2v-720p.safetensors (293MB)
        └── wan21-action-orgasm-v2-t2v-14b.safetensors (585MB)

Total Repository Size: 3.2GB (8 action-specific LoRA adapters)

🎯 LoRA Categories

Camera Angle Adapters

• POV (Point of View): First-person perspective variations
• Front: Frontal camera angles with specific framing
• Dynamic: Movement-following camera tracking

Action-Specific Adapters

• v1: Initial training version with standard settings
• v2: Improved training with enhanced quality and motion consistency

Resolution Variants

• T2V: Text-to-video optimized (standard resolution)
• I2V: Image-to-video optimized (480p/720p)
• 720p: Higher resolution variants for improved detail

🚀 Usage Examples

Basic I2V Usage with Action LoRA

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

# Load base WAN 2.1 I2V model
pipe = DiffusionPipeline.from_pretrained(
    "path/to/wan21-i2v-14b",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load WAN 2.1 VAE
pipe.vae = AutoencoderKL.from_single_file(
    "path/to/wan21-vae.safetensors"
)

# Load action-specific LoRA with absolute path
pipe.load_lora_weights(
    "E:/huggingface/wan21-fp8-loras-nsfw/loras/wan/wan21-action-missionary-pov-i2v-v2.safetensors"
)

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

# Generate video with action-specific motion
prompt = "smooth camera movement, cinematic lighting, high quality"
video = pipe(
    prompt=prompt,
    image=input_image,
    num_inference_steps=50,
    guidance_scale=7.5,
    num_frames=24
).frames

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

Switching Between Action LoRAs

# Define base path
LORA_PATH = "E:/huggingface/wan21-fp8-loras-nsfw/loras/wan"

# Available action LoRAs
actions = {
    "bounce_v1": "wan21-action-bounce-v1.safetensors",
    "doggy_pov_v2": "wan21-action-doggypov-v2.safetensors",
    "doggy_front_t2v": "wan21-action-doggystylefromfront-t2v.safetensors",
    "front_doggy_plow_v2": "wan21-action-front-doggy-plow-v2.safetensors",
    "missionary_pov_i2v_v2": "wan21-action-missionary-pov-i2v-v2.safetensors",
    "missionary_pov_v2": "wan21-action-missionary-pov-v2.safetensors",
    "orgasm_i2v_720p": "wan21-action-orgasm-i2v-720p.safetensors",
    "orgasm_v2_t2v": "wan21-action-orgasm-v2-t2v-14b.safetensors",
}

# Load specific action
selected_action = "missionary_pov_v2"
pipe.load_lora_weights(f"{LORA_PATH}/{actions[selected_action]}")

T2V Usage (Text-to-Video)

# Load T2V action LoRA
pipe.load_lora_weights(
    "E:/huggingface/wan21-fp8-loras-nsfw/loras/wan/wan21-action-orgasm-v2-t2v-14b.safetensors"
)

# Generate from text prompt
prompt = "cinematic scene with dynamic movement, professional lighting, high detail"
video = pipe(
    prompt=prompt,
    num_inference_steps=50,
    guidance_scale=7.5,
    num_frames=24,
    height=720,
    width=1280
).frames

⚙️ Technical Specifications

Model Architecture

• Type: LoRA (Low-Rank Adaptation) adapters
• Precision: BF16 (Brain Floating Point 16-bit)
• Format: SafeTensors (.safetensors)
• Base Compatibility: WAN 2.1 I2V 14B architecture
• Training Method: Action-specific motion patterns and camera movements

LoRA Variants

v1 LoRAs (Initial Generation):

• File sizes: 293-343MB
• Standard motion patterns
• Baseline quality and consistency

v2 LoRAs (Enhanced Generation):

• File sizes: 293-586MB
• Improved motion consistency and temporal coherence
• Better training datasets and methodology
• Recommended for production use

Resolution Support

I2V 720p Models:

• Input: 1280x720 images
• Output: High-resolution video sequences (720p)
• Higher VRAM requirements (20GB+)
• Best visual quality

Standard I2V Models:

• Input: 480p or 720p images
• Output: Matching resolution video
• Balanced quality and performance (16-17GB VRAM)

T2V Models:

• No input image required
• Resolution specified in generation parameters
• More flexible but requires better prompting

💻 Hardware Requirements

Minimum Requirements

• GPU: NVIDIA RTX 3060 (12GB VRAM) or AMD equivalent
• RAM: 16GB system RAM
• Storage: 4GB for LoRAs + base model space (~50GB total)
• GPU Architecture: Ampere or newer (BF16 support required)

Recommended Configuration

• GPU: NVIDIA RTX 3090 (24GB VRAM) or RTX 4070 Ti (16GB)
• RAM: 32GB system RAM
• Storage: 100GB SSD (for full WAN 2.1 ecosystem)
• Operating System: Linux or Windows 10/11

High-End (720p Generation)

• GPU: NVIDIA RTX 4090 (24GB VRAM) or A100 (40GB)
• RAM: 64GB system RAM
• Storage: 200GB+ NVMe SSD
• Optimized for: 720p high-quality output at 50+ inference steps

Software Requirements

• Python: 3.9+ (3.10 or 3.11 recommended)
• PyTorch: 2.0+ with CUDA 11.8 or 12.1
• Diffusers: 0.25.0 or newer
• Transformers: 4.36.0 or newer
• CUDA Toolkit: 11.8+ or 12.1+

Installation

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

# Install diffusion libraries
pip install diffusers transformers accelerate safetensors

# Install image/video utilities
pip install pillow opencv-python imageio

📊 Performance Benchmarks

I2V Generation Speed (NVIDIA RTX 4090, 24 frames)

Hardware Comparison (720p, 50 steps, 24 frames)

Note: Actual performance varies based on prompt complexity, frame count, and system configuration.

🎨 Prompting Tips

Effective Prompts for Action LoRAs

Camera and Movement Descriptors:

• Camera angles: "POV perspective", "frontal view", "tracking camera", "dynamic angle"
• Motion quality: "smooth movement", "dynamic motion", "fluid transitions", "steady camera"
• Lighting: "cinematic lighting", "soft lighting", "dramatic shadows", "natural lighting"

Quality Modifiers:

• Resolution: "high quality", "detailed", "HD quality", "720p quality"
• Style: "realistic", "photorealistic", "cinematic style", "professional"
• Atmosphere: "cinematic", "dramatic", "intimate", "intense"

Example Prompts by LoRA Type:

# POV LoRAs (missionary-pov, doggy-pov)
"POV perspective, smooth camera movement, cinematic lighting, high quality, realistic"

# Frontal LoRAs (front-doggy-plow, doggystylefromfront)
"Frontal view, dynamic motion, professional lighting, detailed composition, 720p quality"

# 720p LoRAs (orgasm-i2v-720p)
"Tracking camera, fluid motion, soft lighting, high detail, photorealistic, HD quality"

# T2V LoRAs (orgasm-v2-t2v-14b, doggystylefromfront-t2v)
"Cinematic scene, dynamic movement, professional lighting, realistic, high detail, 720p"

Action-Specific Guidance

• POV LoRAs: Emphasize first-person perspective and immersive camera work
• Frontal LoRAs: Focus on frontal framing, composition, and balanced lighting
• 720p LoRAs: Include high-resolution quality modifiers and detail keywords
• T2V LoRAs: Provide more detailed scene descriptions (no input image guidance)
• Bounce LoRAs: Mention rhythm, movement dynamics, and motion fluidity

🔧 Troubleshooting

Out of Memory (OOM) Errors

# Solution 1: Enable model CPU offloading
pipe.enable_model_cpu_offload()

# Solution 2: Reduce frame count
video = pipe(prompt, num_frames=16, image=input_image)  # Instead of 24

# Solution 3: Lower resolution
video = pipe(prompt, height=480, width=854, image=input_image)

# Solution 4: Sequential CPU offload (for extreme constraints)
pipe.enable_sequential_cpu_offload()

# Solution 5: Use attention slicing
pipe.enable_attention_slicing(slice_size="auto")

Poor Motion Quality

Causes and Solutions:

• Use v2 LoRAs instead of v1 for better temporal consistency
• Increase inference steps to 40-60 (vs 20-30 for distilled models)
• Adjust CFG scale to 7.0-8.5 range for action LoRAs
• Ensure high-quality input images (for I2V models)
• Match frame count to action complexity (24+ frames for complex actions)

Inconsistent Action Output

Optimization Strategies:

• LoRA Selection: Choose action-specific LoRA matching your use case
• Prompt Alignment: Ensure prompts match LoRA's trained patterns
• Guidance Scale: Use higher guidance (7.5-8.5) for controlled actions
• Input Quality: Higher resolution and quality inputs produce better results
• Frame Count: Some actions require 24+ frames for smooth completion

LoRA Selection Guide

📝 Model Card

📄 License

These LoRA adapters are subject to the WAN model license terms with additional restrictions:

Required Safeguards

• Age Verification: Must implement age verification for end users (18+)
• Legal Compliance: Users responsible for compliance with local laws and regulations
• Content Warnings: All distributions must include appropriate NSFW warnings
• Platform Policies: Respect hosting platform content policies and terms of service

Prohibited Uses

• Non-consensual content generation or deepfakes
• Content featuring or depicting minors
• Identity theft, impersonation, or exploitation
• Harassment, abuse, or harmful content creation
• Violation of individuals' rights or platform terms

Commercial Use

• Check WAN base model license for commercial usage restrictions
• Implement appropriate content moderation systems
• Maintain age verification and legal compliance infrastructure

⚖️ Ethical Guidelines

Prohibited Uses

• ❌ Non-consensual content generation or deepfakes
• ❌ Content featuring minors in any form
• ❌ Identity theft, impersonation, or exploitative materials
• ❌ Harassment, abuse, or harmful content
• ❌ Violation of platform terms of service or community guidelines

Recommended Best Practices

• ✅ Implement robust age verification systems (18+)
• ✅ Use clear content warnings and NSFW tags
• ✅ Respect intellectual property, privacy, and likeness rights
• ✅ Implement content moderation and reporting mechanisms
• ✅ Provide user control and opt-out mechanisms
• ✅ Follow platform-specific content policies
• ✅ Maintain transparency about AI-generated content

🙏 Acknowledgments

• WAN Development Team for the base WAN 2.1 I2V 14B model architecture
• Community Contributors for responsible testing, feedback, and ethical guidance
• Hugging Face for hosting infrastructure with content policy enforcement
• Open-source Community for diffusers, transformers, and safetensors libraries

📚 Related Resources

Related Model Repositories

• WAN 2.1 Base Models: wan21-fp16, wan21-fp8 (I2V base models)
• WAN 2.1 VAE: Required for video decoding and encoding
• WAN 2.1 Camera LoRAs: wan21-camera-* (SFW camera control LoRAs)
• WAN 2.2 Models: wan22-fp16, wan22-fp8 (newer generation models)

Documentation and Tutorials

• Diffusers Documentation: https://huggingface.co/docs/diffusers
• WAN Official Documentation: Check base model repository for guides
• LoRA Training Guide: https://huggingface.co/docs/peft/conceptual_guides/lora

Community Resources

• Hugging Face Forums: Community support and discussions
• GitHub Issues: Technical support and bug reports
• Model Cards: Detailed specifications and usage examples

📧 Support and Contact

Technical Support

• Repository Issues: Open an issue in this repository for technical problems
• Base Model Questions: Refer to WAN official documentation and repositories
• Library Issues: Check diffusers/transformers GitHub repositories

Ethical Concerns

• Content Violations: Report to platform moderators or administrators
• Legal Issues: Consult with legal counsel regarding compliance
• Safety Concerns: Follow platform reporting procedures

Feature Requests

• Community feedback and feature requests are welcome
• Consider contributing improvements or additional LoRAs
• Share responsible usage patterns and best practices

📋 Summary

This repository contains 8 specialized action LoRA adapters for the WAN 2.1 I2V 14B video generation model:

Key Specifications

• Total Size: 3.2GB (8 action-specific adapters)
• Content Type: Adult/NSFW (18+ only)
• Variants: v1 (standard) and v2 (enhanced)
• Resolution: 480p and 720p support
• Precision: BF16 (Brain Floating Point 16-bit)
• Format: SafeTensors (.safetensors)

Use Cases

• Specialized action sequences with controlled motion patterns
• Camera angle control (POV, frontal, tracking)
• Text-to-video and image-to-video generation
• High-resolution video synthesis (up to 720p)

Requirements

• WAN 2.1 I2V 14B base model (required)
• WAN 2.1 VAE (required for video encoding/decoding)
• 12-24GB VRAM (depending on resolution and settings)
• BF16-capable GPU (NVIDIA Ampere/Ada or newer)

Content Warning

These models are trained on adult content and intended for responsible adult use only (18+). Users must:

• Comply with applicable laws and regulations
• Implement appropriate age verification and safeguards
• Use ethically and respect individual rights
• Follow platform content policies and guidelines

Technical Note

These are specialized LoRA adapters that modify the base WAN 2.1 I2V model behavior to generate specific action sequences and camera movements. They require the WAN 2.1 base model and VAE to function and cannot be used independently.

---

Last Updated: October 2025 Repository Version: 1.4 Total Size: 3.2GB (8 action LoRAs with v1/v2 variants) Content Rating: Adult/NSFW (18+ only) Primary Use Case: Specialized action and camera angle video generation for adult content License: WAN model license with additional ethical use restrictions