Wan2.1-Turbo-fp8

This is the ultimate optimized version of Wan-AI's Wan2.1-T2V-14B, combining both Turbo LoRA acceleration and fp8 quantization. This package offers the best balance of speed, memory efficiency, and quality:

• Ultra-Fast: Video generation requires only 4 steps instead of 30-40, leading to ~8-10x speed-up
• Memory Efficient: FP8 quantization reduces VRAM requirements significantly
• High-Quality: Maintains generation quality equivalent to the base model
• Optimized Inference: Combines LoRA acceleration with fp8 for maximum efficiency

Features

• 4-Step Inference: Dramatically reduced sampling steps (30-40 → 4)
• FP8 Quantization: Both model and LoRA weights quantized for reduced memory footprint
• Lower VRAM Requirements: Enables deployment on GPUs with less memory
• Fastest Inference: Combines speed benefits of Turbo LoRA and fp8
• Equivalent Quality: Delivers visuals on par with the full precision 30-40 step base model
• Complete Package: Contains all Wan2.1-T2V-14B components (fp8) + Turbo LoRA weights (fp8)
• Production-Ready: Ideal for applications requiring both speed and efficiency

Usage

This model is planned to work with Aquiles-Image, providing an OpenAI-compatible API for video generation:

pip install aquiles-image
aquiles-image serve --model "wan2.1-turbo-fp8"

Note: Support for this model in Aquiles-Image is coming soon. Stay tuned for updates!

Learn more in the full documentation.

Example of a Video Generated with This Model

Generated with prompt: A direct continuation of the existing shot of a chameleon crawling slowly along a mossy branch. Begin with the chameleon already mid-step, camera tracking right at the same close, eye-level angle. After three seconds, its eyes swivel independently, one pausing to glance toward the lens before it resumes moving forward. Maintain the 100 mm anamorphic lens with shallow depth of field, dappled rainforest light, faint humidity haze, and subtle film grain. The moss texture and background greenery should remain consistent, with the chameleon's deliberate gait flowing naturally as if no cut occurred.

Acknowledgments

We extend our gratitude to:

• The Wan-AI Team for developing and open-sourcing Wan2.1-T2V-14B
• The lightx2v team (@lightx2v) for training the Turbo LoRA and providing the high-quality fp8 quantized weights

This repackaged version is made possible by their excellent work and commitment to the open-source community.

For more information about the original model, please visit the official Wan2.1-T2V-14B repository.