October 2024 Summaries

The Recraft Team has announced the release of the Recraft V3 model, codenamed "red_panda," in partnership with recraft.ai and available on fal. This model has achieved the highest Elo score in the Artificial Analysis image arena, highlighting its advanced capabilities in generating realistic, illustrated, and vectorized images. Users can explore its features, including custom styles and color controls, through fal's demo page. The Recraft V3 model offers a wide range of built-in styles, from 3D to line art, making it a versatile tool for image creation. Interested users can experiment with these functionalities on fal.ai at no extra cost.

Stable Diffusion 3.5, developed by stability.ai in partnership with fal, introduces new models designed for enhanced customizability, efficient performance, and diverse image outputs. The release includes the Large model with 8 billion parameters for superior image quality and prompt accuracy suitable for professional use, and the Turbo model, a streamlined version offering fast, high-quality image generation in just four steps. Both models run efficiently on standard consumer hardware. Additionally, a Medium variant, featuring 2.5 billion parameters and improved architecture, is set to release on October 29th. These advancements aim to provide users with flexible, high-performance tools for creative and professional applications.

Training a style LoRA on the Fal platform involves several crucial steps and considerations to effectively capture and reproduce artistic styles using the FLUX model. Central to this process is compiling a high-quality dataset of images that accurately represent the desired style, with a preference for high-resolution images, such as those of the 19th-century painter Thomas Cole. The number of images needed depends on their quality and consistency, as a smaller set of high-quality images is often more effective than a larger set of lower quality. Captioning plays a vital role in associating the style with specific prompts, where a unique trigger phrase can enhance style reproduction while maintaining prompt flexibility. The step count during training is also essential, as it affects the model's ability to retain the style without losing creative prompt-following capabilities. Custom captioning with both short and long captions can further refine results, allowing for better style and content separation. Experimenting with different step counts and captioning methods can optimize the training outcomes, leading to more precise style transfer and creative expression in generated images.

Fal has collaborated with Black Forest Labs to introduce the FLUX1.1 [pro] model, also known as "blueberry," and enhance the speed of the FLUX.1 [dev] models. The FLUX1.1 [pro] model, which was recently showcased in the Artificial Analysis image arena, has achieved the highest Elo score among image models while maintaining impressive speed. Additionally, the FLUX.1 [dev] model has set new speed records for LoRA training, completing 1000 steps in under two minutes, making it the fastest by a significant margin. A new Reference API endpoint for the FLUX.1 [dev] model has been launched, doubling the speed of the existing endpoint without sacrificing image quality, and it is currently available in an experimental version with plans for broader release.