Today, we are announcing Recraft V3, our latest model that sets a new quality standard in the image generation space, outperforming all competitor models proven by the Hugging Face’s industry-leading Text-to-Image Benchmark by Artificial Analysis.

The new Recraft V3 delivers across-the-board improvements, with particularly notable advances in text generation.

We are also launching several new important features that allow our users to have more control over AI generation: the possibility to specify text size and positions in the image, precise style control, improved inpainting, and new outpainting capabilities.

The model is now available for both free and paid users in the desktop app on Canvas, in the mobile app (available on iOS and Android), and via API. 
‍

‍Winning image generation quality benchmarks

Recraft V3 has set a new standard for excellence in image generation. Over the past 4 days Recraft V3 participated in the Hugging Face’s industry-leading Text-to-Image Model Leaderboard by Artificial Analysis. It secured #1 place with ELO rating of 1172. Recraft's new model is showing quality higher than models of Midjourney, OpenAI, and all other major image generation companies. 

‍

Artificial Analysis’s leaderboards are the most popular publicly available rankings that evaluate and compare the performance of machine learning models across various tasks and datasets. Each leaderboard is designed around a specific task or benchmark dataset, showcasing how different models measure performance metrics.

The text-to-image leaderboard ranks image generation models by their ELO scores, determined by pairwise image comparisons labeled by image arena visitors. Originally developed for ranking chess players, the ELO rating system has since been adapted to evaluate models across various competitive benchmarks.

The main advantages of Recraft V3 light in text generation quality, anatomical accuracy, prompt understanding, and high aesthetic quality.

Recraft V3 is the only model in the world that can generate images with long texts, as opposed to just one or a couple of words.

‍

Anatomical correctness is the metric used to select the model that generates the most accurate anatomy, ensuring a proper number of fingers, hands, and legs, realistic body proportions, spatial coherence within the scene, and natural positioning of background objects relative to the main subject. Recraft V3 is tuned to generate images with correct anatomy.

‍

Prompt-following refers to how accurately the image aligns with the details specified in the prompt. Recraft V3 can generate images with complex scenes, including the correct count, color, and positions of objects mentioned in the prompt.

‍

Aesthetical value is a subjective metric of the "beautifulness" of the image. This is where Midjourney has historically shined. The new model Recraft V3 takes this metric into account and is trained to have high aesthetic value for generated images.

State-of-the-art in generation meets control and design capabilities

Recraft started with a goal to build foundation models in the image generation space that would solve the needs of professional designers. The text-to-image benchmark focuses on overall image generation quality. However, for real-world tasks in the graphic design field, having a high-quality text-to-image model is not enough. It is important to provide users full control over image generation so that they can implement their ideas with high precision.

The new model Recraft V3 is trained to provide more control over image generation than all other existing AI models. This release's main innovations are:

‍

Positioning Control and Generation of Graphic designs

The new model allows specifying the exact positions and sizes of text on a design.

‍
It is also possible to position other images and combine them with texts, allowing the generation of complex graphic designs.

‍

Style control 

The improved style creation process allows for fine-grained experimentation. It is possible to select a set of images to represent the brand style, and experiment with the style candidate until it is tuned to the exact look and feel needed for the brand. This is possible because Recraft V3 accepts style as an input to the model, and doesn't require retraining of the model to catch details of the style.

‍

Apart from the improved control features, Recraft V3 supports unique capabilities vital for graphic design space. The distinctive feature of Recraft is that it supports vector image generation, ranging from sets of simplistic pictograms all the way to highly detailed vector art. Also, Recraft V3 provides a whole suite of AI image editing tools that helps designers create and edit visuals end-to-end: AI Eraser, Modify Area, Inpainting, Outpainting, AI Mockuper, Creative and Clarity Upscalers, AI Fine-Tuning, and Background Remover. 

‍The first API with support of style consistency and vector art generation

Recraft has also launched an API that enables developers and businesses to integrate SOTA image generation and AI design capabilities into their workflows. This API provides access to the Recraft model, supporting both raster and vector formats, generation of images with text and allowing for the creation of custom styles to ensure brand consistency. Additionally, it supports specifying brand colors and offers advanced features like vectorization, upscaling, image quality improvement, and background removal, providing users with a unique set of features covering all AI image editing suite via API.

What’s next for Recraft? 

Recraft is all about giving designers more control. This means creators can manage every detail of their designs, ensuring the final product looks exactly how they want it. Recraft's team believes in providing intuitive solutions that allow pro designers to focus on what they do best: creating.

Recrafters are always invited to give us feedback and feature requests, telling us what they love or dislike about the Recraft experience.