MODEL_CARD.md

---

license: apache-2.0 library_name: diffusers tags:

• text-to-video
• image-to-video
• video-generation
• diffusers pipeline_tag: text-to-video base_model: deathlegionteam/LEGION-Video-Gen inference: true widget:
• text: "A serene mountain lake at sunset with colorful clouds reflecting on the water"

⚔️ LEGION Video Generation

State-of-the-art video generation with 8.3B parameters

Model Description

LEGION Video Generation is a production-ready video generation system with 8.3 billion parameters. It supports text-to-video (T2V) and image-to-video (I2V) generation with temporal enhancement and configurable QWatermark system.

Architecture

Component	Description
Transformer	54-layer 3D Diffusion Transformer
VAE	3D causal VAE with 32-channel latent space
Text Encoder	Qwen2.5-VL (7B) + T5 Encoder ensemble for rich text understanding
Scheduler	Flow Matching Euler Discrete Scheduler with shifting
Parameters	8.3 Billion total
Precision	FP16 (inference)

Key Features

• 🎬 Text-to-Video — Generate cinematic videos from any text prompt
• 🖼️ Image-to-Video — Animate still images with controlled motion
• 💧 QWatermark System — Configurable quality assurance watermark overlay
• 🌐 Web UI — Gradio frontend with dark theme and FastAPI backend
• 📡 REST API — Programmatic access via HTTP endpoints

Intended Use

Direct Use

• Video Generation: Create high-quality videos from text descriptions
• Content Creation: Generate video assets for social media, marketing, and creative projects
• Animation: Animate still images with natural motion
• Prototyping: Rapid video ideation for film, game, and design workflows

Out-of-Scope Use

• Generating deceptive or misleading video content (deepfakes)
• Creating violent, hateful, or otherwise harmful content
• Misrepresenting generated content as authentic footage
• Bypassing content safety systems

How to Get Started

Installation

# Clone the repository
git clone https://huggingface.co/deathlegionteam/LEGION-Video-Gen
cd LEGION-Video-Gen

# Install dependencies
pip install -r requirements.txt

Basic Inference

from diffusers import DiffusionPipeline
import torch

# Load the pipeline
pipe = DiffusionPipeline.from_pretrained(
    "deathlegionteam/LEGION-Video-Gen",
    torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")

# Enable memory optimizations
pipe.vae.enable_tiling()
pipe.enable_attention_slicing()
pipe.enable_model_cpu_offload()

# Generate a video
video_frames = pipe(
    prompt="A serene mountain lake at sunset with colorful clouds reflecting on the water, cinematic quality",
    negative_prompt="warped, distorted, flickering, jittery, low quality, blurry, artifacts",
    num_frames=49,
    width=480,
    height=480,
    num_inference_steps=50,
    guidance_scale=6.0,
).frames[0]

# Save as MP4
import imageio
imageio.mimsave("output.mp4", video_frames, fps=16, codec="libx264")

Using the LEGION Generator Wrapper

from inference import LegionVideoGenerator

generator = LegionVideoGenerator()
video_path = generator.generate_from_text(
    prompt="A cyberpunk city street at night with neon lights reflecting on wet pavement",
    num_frames=49,
    width=480,
    height=480,
    num_inference_steps=50,
    guidance_scale=6.0,
    watermark_strength=0.3,
)
print(f"Video saved to: {video_path}")

QWatermark System

The QWatermark (Quality Watermark) system imprints a configurable assurance marker:

Parameter	Description	Default
Text	Watermark text	"LEGION"
Position	Placement on frame	bottom-right
Font Size	Text size	36
Opacity	Transparency	0.3
Strength	Overall intensity	0.0 (disabled) - 1.0 (full)

Limitations

1. GPU Required: Real inference requires 16GB+ VRAM GPU. CPU fallback provides mock/test patterns only.
2. Resolution: Optimized for 480p. Higher resolutions (720p+) require more VRAM.
3. Video Length: Generates up to 129 frames (~8 seconds at 16 FPS).
4. Content Quality: Results vary with prompt quality; complex scenes may show artifacts.

License

This model is released under Apache 2.0 License.

Contact

• Organization: deathlegionteam/LEGION-Video-Gen

⚔️ LEGION VIDEO GENERATION