🎬 Consumer GPU Video Generation

Run 67GB video models on 32GB consumer GPUs

NVIDIA-style optimizations without ComfyUI dependency

---

🎯 The Problem

Model	Original VRAM	Your GPU
LTX-2 19B	67 GB	❌ Won't fit
Wan 2.2 5B	25 GB	⚠️ Barely fits

Large video generation models require datacenter GPUs. Consumer cards like RTX 4090/5090 can't run them... until now.

---

✨ The Solution

Before (Original) ❌ LTX-2: 67 GB VRAM ❌ Wan 2.2: 25 GB VRAM

After (INT4 Optimized) ✅ LTX-2: 22 GB VRAM (-67%) ✅ Wan 2.2: 16 GB VRAM (-36%)

75% VRAM reduction via INT4 quantization with minimal quality loss.

---

🚀 Key Features

💾 75% Less VRAM INT4 quantization shrinks models to fit consumer GPUs

🔓 No ComfyUI Standalone Python - use in any project

⚡ RTX 5090 Ready Tested on latest Blackwell architecture

🎨 Simple API 3 lines of code to generate video

---

📊 Benchmarks

Tested on NVIDIA RTX 5090 (32GB) with CUDA 12.8

Model	Original	Optimized	Resolution	Speed
Wan 2.2 TI2V-5B	25 GB	16 GB	1280×704	~50s
LTX-2 19B	67 GB	22 GB	640×448	~60s

📈 Detailed VRAM Breakdown

Wan 2.2 TI2V-5B

Component	Original	INT4
T5 Text Encoder	11 GB	11 GB
VAE	3 GB	3 GB
DiT Transformer	11 GB	3 GB
Peak	25 GB	16 GB

LTX-2 19B

Component	Original	INT4
Gemma-3 Text Encoder	27 GB	8 GB
Transformer	40 GB	10 GB
VAE + Audio	5 GB	5 GB
Peak	67 GB	22 GB

---

🛠️ Installation

Requirements

• GPU: RTX 4090, RTX 5090, A6000, or similar (24-32GB VRAM)
• CUDA: 12.0+
• Python: 3.10+

Quick Start

# Clone repository
git clone https://github.com/lumi-node/consumer-gpu-video-gen
cd consumer-gpu-video-gen

# Install dependencies
pip install -r requirements.txt

# Download Wan 2.2 (recommended for most users)
git clone https://github.com/Wan-Video/Wan2.2
huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B

---

💻 Usage

Command Line

# Generate with Wan 2.2
python generate.py --model wan22 \
    --prompt "A fluffy cat walking through a sunny garden" \
    --checkpoint ./Wan2.2-TI2V-5B \
    --wan-repo ./Wan2.2

Python API

from models.wan22 import Wan22Pipeline

# Load with INT4 optimization
pipeline = Wan22Pipeline(checkpoint_dir="./Wan2.2-TI2V-5B", wan_repo_path="./Wan2.2")
pipeline.load(quantization="int4")

# Generate video
video = pipeline.generate("A cat playing in a garden")
pipeline.save_video(video, "output.mp4")

📋 All CLI Options

--model, -m      Model: wan22 or ltx2 (required)
--checkpoint, -c Path to model checkpoint (required)
--wan-repo       Path to Wan2.2 repo (required for wan22)
--prompt, -p     Text prompt (required)
--output, -o     Output path (default: auto-generated)

--frames         Number of frames (default: 33)
--steps          Diffusion steps (default: 30)
--guidance       Guidance scale (default: 5.0)
--seed           Random seed (default: random)
--size           landscape or portrait (default: landscape)
--fps            Output FPS (default: 24)

--quantization   int4, int8, or none (default: int4)

---

🔬 How It Works

┌─────────────────────────────────────────────────────────────┐
│                    Standard Loading                          │
│  ┌─────────┐  ┌─────────┐  ┌─────────────┐                  │
│  │ T5 Enc  │ +│   VAE   │ +│ Transformer │ = 67GB ❌        │
│  │  27GB   │  │   5GB   │  │    40GB     │                  │
│  └─────────┘  └─────────┘  └─────────────┘                  │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│                  INT4 Quantized Loading                      │
│  ┌─────────┐  ┌─────────┐  ┌─────────────┐                  │
│  │ T5 Enc  │ +│   VAE   │ +│ Transformer │ = 22GB ✅        │
│  │   8GB   │  │   5GB   │  │    10GB     │   (INT4)         │
│  └─────────┘  └─────────┘  └─────────────┘                  │
└─────────────────────────────────────────────────────────────┘

The Magic: quanto INT4 Quantization

• 16-bit → 4-bit weights = 75% smaller
• No retraining required - post-training quantization
• Minimal quality loss - optimized dequantization at inference

Smart Memory Management

1. Load models sequentially
2. Quantize before moving to GPU
3. Offload unused models during VAE decode
4. Strategic garbage collection

---

🎮 GPU Compatibility

GPU	VRAM	Wan 2.2	LTX-2
RTX 5090	32 GB	✅ Full	✅ Reduced res
RTX 4090	24 GB	✅ Full	⚠️ Tight
RTX 4080	16 GB	⚠️ Limited	❌
RTX 3090	24 GB	✅ Full	⚠️ Tight
A6000	48 GB	✅ Full	✅ Full

---

🤝 Contributing

Contributions welcome! Areas of interest:

• Additional model support (CogVideoX, etc.)
• FP8 quantization for Blackwell GPUs
• Web UI interface
• Audio generation for LTX-2

---

📚 Acknowledgments

• Alibaba Wan Team - Wan 2.2 model
• Lightricks - LTX-2 model
• Hugging Face - quanto quantization
• NVIDIA - Optimization inspiration from ComfyUI implementations

---

📄 License

MIT License - see LICENSE file.

---

_{Built with ❤️ for the open-source AI community}