WAN 2.5 TEXT TO IMAGE

Wan 2.5 Text to Image is a text-to-image generation model developed by fal.ai. It creates high-quality images from descriptive text prompts provided by the user. The model accepts text input in both Chinese and English, supporting prompts up to 2000 characters in length. Users can specify detailed scene descriptions, artistic styles, or desired visual atmospheres in natural language, and the model generates images closely aligned with these specifications.

Users have advanced control over image generation through a variety of configurable parameters. The 'image_size' parameter allows users to select pre-set standard aspect ratios such as 'square', 'landscape_16_9', or 'portrait_4_3', or to manually define custom dimensions (height and width). The model enforces a pixel count range, allowing images from 768×768 up to 1440×1440, with an aspect ratio spanning from 1:4 to 4:1. This flexibility accommodates a wide range of creative needs, from square compositions to cinematic panoramas or vertical portraits.

For refinement and control, users may include a 'negative_prompt' (up to 500 characters) to identify elements or qualities they wish to avoid in the output (for example: 'low resolution, error, worst quality, low quality, defects'). The model can generate between 1 and 4 images per request, aiding in variation or ideation. To enable repeatable and reproducible results, an optional 'seed' parameter allows users to control random initialization; if a seed is not specified, the process is randomized.

Two significant features provide further customization and safety. The 'enable_prompt_expansion' option, enabled by default, leverages large language models to rewrite or enhance shorter prompts, improving the detail and quality of resulting images (at the cost of increased processing time). The 'enable_safety_checker', also enabled by default, activates a safety mechanism intended to filter outputs, though specific details of safety implementation are not provided.

Outputs are delivered in both image (PNG format) and JSON formats. The JSON response includes the generated image URL, content type, seed used, and the final prompt sent to the model, supporting integration with various workflows and reproducibility. This makes the model suitable for both interactive use via web interfaces (such as the fal.ai Playground) and programmatic access via API endpoints.

An example prompt provided in the documentation demonstrates the model's ability to render complex scenes with dramatic lighting and atmospheric depth: 'A lone samurai standing on the edge of a cliff at twilight, overlooking a vast valley shrouded in mist. The sky burns with deep orange and purple hues from the setting sun, casting long, dramatic shadows. The samurai’s silhouette glows against the horizon, with their sword reflecting a glint of fading light. The overall style is hyper-realistic, cinematic, and moody, with dramatic contrast and atmospheric depth.'

The documentation indicates that Wan 2.5 Text to Image supports commercial use, making it appropriate for partners and business applications in addition to individual creative projects. However, no explicit information about supported industries or user roles is provided.

There are some operational boundaries: image size or aspect ratio constraints must be observed, and prompt, negative prompt, and number of images per request are limited by maximum character count or allowed range. Specific details about processing speed, hardware requirements, or limitations on style or subject matter beyond those handled by the safety checker are not included in the documentation.

Overall, Wan 2.5 Text to Image provides users with a versatile, controllable engine for generating images from textual input, supporting a range of use cases from ideation and design to illustration and commercial content creation, wherever high flexibility and prompt-based customization are valuable.