Start the Flux.1 Capacitor
In the ever-evolving world of AI image generation, a new contender has emerged: Flux.1 developed by Black Forest Labs. With its claim to be on a par with if not better than MidJourney in certain aspects, Flux.1 has attracted the interest of AI enthusiasts and professionals alike.
Flux.1 was created by a team who previously contributed to building Stable Diffusion. This includes the creation of VQGAN and Latent Diffusion--not to mention AI models like Stable Diffusion XL and Rectified Flow Transformers that can generate images or videos. This team will bring cutting-edge experience and innovation to the table.
Flux.1 comes in three versions, each with varying levels of power and cost:
Flux.1 Schnell: The fastest and most widely usable of the systems. It is designed for local development and personal use, being open source under Apache 2.0 license which means it can be used anywhere—commercial or not.
Flux.1 Dev This model bests the Schnell model with its performance and prompt compliance yet less so costs but must be reserved exclusively for non-commercial applications.
Flux.1 Pro: This top model suits enterprise applications and provides truly first-class performance.
Several platforms have already implemented Flux.1, offering free use:
Hugging Face Spaces: Provides the Schnell and Dev models. Users can enter their prompts in this as well as adjust a few settings such as seed, width, height, and number of inference steps in order to generate images.
Glyph With this AI workflow builder using the Pro model for free, users can refine their prompts via LLMs like Claude or ChatGPT and generate high-quality images.
Here are some findings from comparing Flux.1 with other AI tools like MidJourney and DALL-E 3:
Illustrations: MidJourney produced better hand-drawn, oil, and watercolor illustrations than Flux.1 (which appears to have outsourced these tasks while soldering chips on customers' page). The art by Flux.1 generally had a funkier look, unique but curiously off sometimes. You saw similarly (but not quite as well done in our judgment) attempts at Liao's paintings that Dall-W did in 100 lines with PoemMatic.
Realism: Flux.1 is good at realism; for example, when you ask it to create a photo of a man eating ice cream or a woman taking selfies on the beach, it gives impressive results. But MidJourney still comes in slightly ahead on hyper-realism.
Text in Images: Here Flux.1 really shines, delivering clear and beautiful texts involved in images (see The sign—a wide polar bear was holding with specific text or The plane writing in the sky). This column tested Stable Diffusion and found that while it could do some text oddly could not lay out clearly repeating passages such as those required for parodies of Guangzhou rhyme-writing style; even DALL-E 3 did not entirely resolve such problems.Foundation and Dall-e 3
Prompt Adherence:: While Flux.1 generally captures most elements of complex prompts it sometimes fails to do so as well in comparison to DALL-E 3 (which excels in this area). For example, a prompt which involved a three-headed dragon and some specific actions was better handled by DALL-E 3.
Flux.1 is a powerful and promising tool in the AI image generation landscape. While it may not yet surpass MidJourney in every aspect, it offers significant advantages, particularly in text generation and open-source flexibility. As it continues to evolve, Flux.1 has the potential to become a formidable competitor, combining the strengths of MidJourney, DALL-E 3, and Stable Diffusion.
For those looking to explore the latest advancements in AI art, Flux.1 is definitely worth a try. Keep an eye on this tool as it develops and integrates into more platforms, and don't hesitate to experiment with its capabilities. Stay tuned for more updates and insights as the exploration of the exciting world of AI image generation continues!
