A few days ago, in the text-to-image model arena, a new model suddenly emerged and outperformed MJ, Flux, and SD to reach the top globally. The data comes from the Artificial Analysis arena. This model called “red_panda” achieved a 79% winning rate in the human blind test, overpowering top models such as Flux, Midjourney, and Stable Diffusion. Its ELO score is over 100 higher than that of the newly crowned Flux 1.
1 Pro model, and nearly 200 points higher than models like Midjourney 6.1 and Stable Diffusion 3.5 Turbo. It’s incredibly powerful. You know, this was a blind selection by the public, and users didn’t know the specific model manufacturers. The blind selection process is like this. SD VS red_panda: The arena randomly selects two anonymous models for a PK, and users choose the image that better matches the prompt according to their personal aesthetics. In the blind test, the red_panda model achieved a 79% winning rate, and its score reached a record high of 1244. The arena evaluation address: https://artificialanalysis.ai/text-to-image/arena?tab=LeaderboardMystery Revealed
After the Red_panda model became the top-ranked, it became popular overnight, but no manufacturer claimed it for a long time. So, everyone started a guessing game to speculate about its origin. Some said it was OpenAI’s Dall-e4, or Midjourney V7. Others guessed it was Meta… Some even guessed it might be from a Chinese manufacturer. Because red panda, the red panda, has a strong Chinese flavor. Today, the mystery is officially revealed. It doesn’t come from OpenAI, nor from Midjourney, Stable Diffusion, Flux, and it’s not from a Chinese manufacturer either. Instead, it comes from the “Recraft V3” model (see the experience link at the end of the article), developed by the AI startup Recraft. Model trial address: https://fal.ai/models/fal-ai/recraft-v3 Recraft is headquartered in London, UK. It’s a startup that has been established for less than two years. They are committed to using AI technology to help designers build the entire design process.Recraft V3 is now available for experience and use on the web version, functioning somewhat like a web-based Photoshop, but with integrated AI capabilities. Compared to other text-to-image models, Recraft V3 has made significant strides: 1) Support for long text generation. Recraft V3 can understand and generate long-form text content, making it the only global text-to-image model capable of producing lengthy texts rather than just a word or a few words. Test case example: Prompt – ‘At night, a sports car is speeding on a racetrack, with the car logo “BYD” and a huge sign reading “China”.’ 2) Controllable text size and positioning. In addition to generating long texts, Recraft V3 can also precisely control the size and positioning of the text. Prompt examples: ‘Design a cool logo with the text “WoYin”, on a pure black background with a gradient purple-blue font, exuding a sense of technology.’ ‘Poster, with the large text “News” at the top, followed by “Recraft V3 is the only model in the world that can generate images with long texts.”’ 3) Optimization of body part completeness. Recraft V3 has significantly optimized the completeness of body parts, ensuring the correct number of fingers, hands, and legs in generated human images, realistic body proportions, spatial consistency in the scene, and natural positioning of background objects relative to the subject. Currently, many AIs inexplicably add extra fingers or have disproportionate height, hand, and leg ratios, making them easily identifiable as AI creations. Midjourney has also been focusing on optimizing body part completeness, but Recraft has taken the lead. Here is my test case, which is very close to a real-life photo. Prompt: ‘A man and a woman sitting side-by-side at a circular dinner table. On the left is an older white man with a gray mustache wearing a light blue button-down shirt. On the right is a younger white woman in a black sequin cocktail dress. ‘
They are in a dimly lit fine-dining restaurant with crystal chandeliers and waiters wearing tuxedos. The older man is eating a green salad, and the younger woman is eating a large steak. Precise Style Control. Recraft V3 accepts style as a model input without the need for retraining to capture details. Simply select a set of images to represent VI aesthetics and refine the candidate styles until the generated images perfectly match the desired appearance and feel. Recraft offers 24 styles that can be used as model inputs. In addition to this, Recraft also supports layered editing, vector processing, API calls, and provides a rich online editing area. How to Use? The Recraft V3 model is now officially online on the Recraft.ai website. Official website address: https://www.recraft.ai. No magic required, register with a QQ email, and you can use it directly. Free users receive 50 credits per day, allowing the generation of 50 images. If you are a new user, by entering through the following invitation link, you can also get an additional 200 credits (200 images) for free. Invitation link: https://www.recraft.ai/invite/5C9e7Hq6ih. Recraft also supports Chinese prompt words, so you no longer need to remember those awkward English ‘spells’. No magic, no Google account, no money, no English prompt words, no need to deploy locally… and it’s even better than Midjourney, Flux, and Stable Diffusion, so try it out now! Oh, by the way, Recraft’s generation speed is also super fast, able to generate 2 images in 14 seconds, which is really impressive. Below are more than 10 selected cases from an entire afternoon of testing. What do you think? With over 100 tests completed, I’ve nearly exhausted 250 credits. Please like, watch, and share my post, it’s not too much to ask, right? T~T Prompt: Lego Halloween assembly, master level, high-end photography, light and shadow, solid color background, high resolution, background blur.
A Q-version wool felt doll presented in 3D, dressed in a cute and mysterious wizard costume. A girl swimming underwater with a Katsuhiro Otomo style and hyper detailed render style. Zhu Shan’s photography and illustrations, with a grainy texture, focusing on feminine beauty, with red, black, and white as the main colors, featuring a close-up of the face with a strong contrast of colors. An anime poster of a woman standing in a city, reminiscent of 1990s anime, with citypop, Showa girl, and retroism style. A beautiful woman with black hair posing in a lace dress. A photo of a 90’s desktop computer on a work desk, with the screen displaying ‘hello’. In the background on the wall, there is beautiful graffiti with the text ‘CHINA’ prominently featured. A studio photograph closeup of a chameleon over a black background. An anime style illustration of a newsstand on top of a small grassy hill, with the text ‘Red Panda’ on the newsstand. In the background, a big rain is approaching. A Cyberpunk cityscape at dawn, with the protagonist standing on a skyscraper rooftop, neon lights fading, futuristic skyscrapers with giant holograms, airships in the distance — all in a cinematic anime style, with rain reflections and vibrant colors.
Discover the enchanting world of miniature landscapes where half a coconut conceals an entire city, complete with ocean, beach, and sun umbrellas. Imagine tiny figures standing amidst a basket of large tomatoes, capturing the essence of miniature photography. Experience a serene view of the sea from an open balcony in Koz, featuring neutral colors accented by pink flowers and green leaves climbing the walls. This article also includes the very first image, generated by AI. Observe an orange-red panda working diligently at a computer, with the word ‘Recraft’ written on the wall. That concludes this article. If you enjoyed this piece, don’t forget to like, watch, and share. Thank you! To receive updates immediately, remember to bookmark us.
