Meet Recraft V3: the best AI image generator you’ve never heard of

Flux and MidJourney aside: there's a new player shot together at AI image generation levels. The mysterious model formerly known as Red Panda—which had AI watchers scratching their heads over artificial intelligence leaderboards—has finally revealed itself as Recraft V3, a new release from a well-known London startup.

The model scored the highest score for image creators on the Flux 1.1 Pro and MidJourney Excellence on the LO rating system. In terms of efficiency, the Recraft V3 matches SDXL's generation speed in less than 10 seconds while delivering superior image quality as blind tests show.

In four days of benchmark testing, Recraft V3 demonstrated excellence in text generation quality, body accuracy, and fast recognition. Capable of generating images with extended text passages, it stands as the only model beyond simple word or phrase synthesis.

In fact, even former stability AI researcher Joe Pena, who worked on SDXL development until Flux appeared, was pretty cool.

“Wow! “I'm very impressed,” said the stunning new model.

Unless you're deep into generative AI or digital design, you probably haven't heard of it, but London startup Recraft AI launched in 2015. It was founded in 2022 and started out as a niche player that focused on AI-powered tools rather than general image generation. Earlier this year, the trajectory changed after securing $11 million in funding from the likes of Khosla Ventures and former GitHub CEO Nat Friedman.

“So it's not fair to say we were unknown, but the buzz around the company was minimal,” Recraft CEO Anna Veronika Dorogush told Decrypt. Just through social media or pure word of mouth.

Recraft V3 excels at creating accurate images, handling fine details and imperfections with noticeable accuracy and works on a subscription model similar to Midjourney, Leonardo or Ideogram.

This Recraft is an evolution of the 20 billion parameter model launched in March – for reference SD 3.5 is a large 8.1b parameters and Flux is a 12 billion parameter model, “parameters” are good as fine-tuned settings developer – tunes during training so that the model understands and responds correctly to complex tasks to do.

“This is a very powerful model, the size of Oder is greater than all open source models. This model was already powerful – according to our internal parameters, it was very modern in anatomical accuracy and fast tracking.” Anna told Decrypt, “But we saw that the old model lacked things – it could not produce correct text. Also, the aesthetic properties of images in the default styles could be improved.”

“We trained the new model there, and now it's beating all the parameters.”

Seven months later, and we have a modern model that has jumped to the best position among all image generators.

Digital design is at the heart of Recraft's values. This model is capable of processing text into vector generations, which means users can request that a model generate images that can be infinitely scaled without losing quality.

Free users receive 50 daily credits which is enough for 50 images. However—and this is a new business model, users do not retain ownership of their inventions. Reserved for paid subscribers, plans starting at $10 per month for 1,000 credits.

How to use Recraft V3

Users can access Recraft V3 through three channels: web interface, Discord commands, or mobile apps available on iOS and Android.

Fire up Discord, join Recraft's server, and if you've ever used Midjourney, you'll find yourself in familiar territory. Go to the #image-gen channel, type “/recraft” and watch the magic happen.

You can also use different modifiers after the question. Looking for a widescreen masterpiece? Throw away –ar 16:9. Need a portrait? Type –ar 3:4, and it has your back.

Other useful additions are the –style command, which allows users to choose unique images for their generation, whether photorealistic, 3D or kawaii. Do you have a different vision in mind? The –sref command lets you upload reference images to guide AI rendering.

After the image is created, users can choose between 2 generations of the image they like the most and then save it or increase the size up to 4 times.

The web interface on recraft.ai flexes some heavy muscles. To get started, simply go to Recraft.AI and log in.

Once on the image generation UI, users only need to enter their query in the left-sized text box.

They also have sliders to change the aspect ratio and number of images.

Users can change the style by clicking on the button on the icon at the top of the text box and choosing from a pop-up menu with several examples.

The interface is more sophisticated than other websites and at first glance it's clear that it's aimed at designers. It allows users to create frames, product mockups, image collections, link to backgrounds, release images, etc.

Mobile creators are not forgotten. Official apps on iOS and Android are available, offering similar quality builds. Simply download the application, log in, click the button in the top center and create an image by pressing the prompt in the text box at the bottom of the screen and clicking the generate button.

Users can choose how detailed the image will be, aspect ratio, styles and references on the same interface. It is very familiar.

Testing the model

We tested the model in several areas: style and technical capabilities. Here's how it stacks up against its competitors—open source and closed source.

Realism

Prompt: A projection of the word “pop” on a woman's face

Recraft V3:

Recraft features a better understanding of natural skin texture, facial contours, and ambient lighting. The projection seems to be well integrated with the skin, and crucially, there are real flaws–Visible pores, minor skin imperfections, and natural hair flyaways. Correct expression and background context add significant accuracy.

Stable Distribution 3.5:

SD 3.5 comes in a close second place. It's a huge improvement over SD3 medium and even the best true SDXL scans. The bright orange color and bright red lipstick show the presence of intense drama. While the facial features are well defined, there is a noticeable artificial quality to the skin texture. The projection looks sharper overlaid, and the expression looks somewhat more laid back and composed.

Midway:

As always, Midjourney creates an emotional, cinematic spectacle with a strong technical performance. However, the woman's skin has a radiant, almost ethereal quality that feels less natural than Recraft's attempt at beauty. The projection is mixed well, but the overall perfection of the features and textures – and the obvious lack of accuracy in the descriptions – shows the origin of the AI.

Recraft is the best model in this comparison.

Quick tracking and spatial awareness

Interviewer: A dog stands on a TV and the word “decrypt” appears on the screen. On the left is a woman in a business suit holding a coin, on the right is a robot standing on a first aid kit. The overall look is realistic.

Recraft V3:

The model was not the best in terms of spatial awareness. However, he managed to achieve a surreal style in the overall composition. This was a departure from other models that showed better compliance and spatial awareness in elements, but the overall feel or style of the scene was questionable.

This may be seen as a good trade-off for some, as it is easier to paint and edit elements within a composition than to re-draw an entire image. But it is important to consider when comparing the Recraft V3 with other models.

Example and metaphor:

Quick: Hand-drawn, creepy, scary, dark and creepy look of a giant spider chasing a woman in the forest, horror, analog photography effects, sketch.

The model has many different styles to choose from, but for this generation we went with Recraft RAW. At first we thought the “hand drawn” style was the best option, but… no, it wasn't.

After trying different presets, good old RAW (the most versatile) was the best fit for what we wanted.

Compared to the other models, Recraft created an interesting composition, and it was accurate in showing the key message of the show: a giant spider chasing a woman. However, the overall art looks more like a digital rendering than a hand-drawn image.

Aesthetically, the most appropriate model for this particular question seems to be the recent Stable Diffusion Model, which created a hand-drawn illustration and managed to convey the stress of a woman running away from a giant spider.

Conclusions

It's easy to see why Recraft V3 ranks high on the image generation leaderboard. Unlike competitors like the MidJourney and Flux, which often have predictable, stylized styles—the bland “Flux face” or the lifeless “MidJourney look”—Recraft V3 leans toward realism. The results are convincing, revealing intricate details such as natural skin texture, subtle imperfections and sub-lights. This aesthetic balance, favoring precision without sacrificing polish, gives Recraft an edge that other models struggle to match.

Pricing strategy is also important to consider. Recraft offers a free tier with generous daily credits, and its support from text to SVG generation is a boon for illustrators looking for scalable, professional-quality vectors. It's about the same price as MidJourney's cheapest plan—but unless you're looking for MidJourney aesthetics, Recraft is more versatile and powerful, so this is a better option.

That said, Recraft isn't perfect. When it comes to complex scenes with multiple elements, spatial awareness sometimes falters. Requests that require precise composition may cause minor discrepancies, and users may paint or edit areas more than expected. But for those who prioritize accuracy and versatility, this shortcoming is easily overlooked.

Also, free users not owning their creations can be a big drawback to consider.

Overall, the Recraft V3 looks like a great close-source option that offers great value and flexibility at a price point that respects a creator's budget. Recraft is a clear winner for anyone looking for high-quality realistic images without the “AI look” trademark.

However, those who can run AI models locally, may be good enough with Flux or SD 3.5.