Wan 2.5 Preview

Wan 2.5 Preview API Overview

Wan 2.5 Preview is the latest iteration of the Wan series text-to-image models, delivering high-fidelity image generation from textual prompts. This release removes previous restrictions on image side length, enabling flexible and unrestricted pixel dimension choices within a defined pixel area. Wan 2.5 combines advanced AI architecture with enhanced pixel-level control to produce diverse and highly detailed visuals.

Technical Specifications

• Model Type: Text-to-Image generative model
• Architecture: Advanced diffusion-based generative network
• Input: Text prompts in natural language
• Output: Variable resolution images, any dimension within supported pixel range
• Training Data: Diverse multimodal dataset including art, photos, and digital illustrations
• Languages Supported: Primarily English, adaptable to other languages with tokenization

Performance Benchmarks

• FID Score (Fréchet Inception Distance): 13.5 on standard image generation benchmarks, indicating high realism and quality.
• Inference Speed: Average generation time of 4 seconds per 512x512 image on modern GPUs.
• Memory Usage: Optimized to run on 12GB and above GPU VRAM configurations.
• Resolution Support: Successfully generates images up to 4K and beyond without quality degradation.
• Diversity: Generates a wide range of unique images for the same prompt, supporting creative exploration.

Wan 2.5 Preview API Pricing 

• $0.0315 / image

Key Features

• High-Quality Detail: Produces crisp and intricate image features across various styles and subject matters.
• Flexible Style Adaptation: Capable of generating artistic, realistic, or stylized images based on prompt context.
• Fast Inference: Efficient model design enables quicker image generation compared to previous versions.
• Scalable Resolution: Suitable for small digital thumbnails up to large-scale prints and presentations.

‍

Code Sample

Comparison with Other Models

vs Stable Diffusion: Wan 2.5 is optimized for high-resolution images with fast inference and consistent quality at large sizes, while Stable Diffusion sometimes experiences quality degradation when scaling up.

vs DALL·E 3: Wan 2.5 Preview provides flexible dimension control enabling users to adapt output sizes freely, making it particularly advantageous for specialized design and print applications.

vs Midjourney: Wan 2.5 Preview is more versatile in dimension customization and supports both stylized and photorealistic outputs with rapid generation, appealing to users needing size flexibility without sacrificing detail.

vs Imagen: Wan 2.5 Preview surpasses Imagen by allowing free selection of image dimensions within pixel area limits, providing more adaptability for diverse use cases and print-ready results.

‍