AI images just got WAY too real. FLUX 1.1 deep dive

The paragraph introduces Flux 1.1 Pro, the latest and best AI image generator available. It mentions that the video will compare Flux 1.1 Pro with other leading image generators based on benchmark scores, but the presenter prefers to focus on practical testing. Various platforms where Flux 1.1 Pro can be used are listed, including Together AI, Rubber Band, Ever Art, Foul, Replicates, and Free.com. The presenter chooses Together AI for demonstration due to the free $5 credits, allowing for several dozen image generations. The process is simple: users can input a positive prompt and adjust image dimensions, but there are no advanced settings like CFG scale or negative prompts. The video demonstrates the generation of images starting with a simple prompt like 'a woman in the city' and discusses the typical AI-generated look with polished faces and blurry backgrounds. A trick to generate more realistic images is introduced: using the prompt 'IMG uncore' followed by any four numbers and 'CR2', which simulates a file structure from a Canon camera, resulting in images that look like they were taken by a regular person rather than AI.

This paragraph delves into the techniques for generating more realistic images using Flux 1.1 Pro. The presenter demonstrates how changing the prompt to include 'CR2' and numbers can produce images that mimic amateur photography, with grainy textures and non-blurry backgrounds. Examples include generating images with prompts like 'cat', 'selfie', and 'party', which result in photos that closely resemble real, non-AI generated images. The paragraph also highlights the limitations of these techniques, noting that they work best with simple prompts and can fail with more detailed descriptions. The presenter experiments with different file formats like 'heic' and different numbers to see variations in the realism of the generated images. The video also shows how prompts like 'selfie.JPEG' can produce highly realistic selfie photos, and the presenter encourages viewers to be creative with their prompts, suggesting that Flux 1.1 Pro can generate images that are nearly indistinguishable from real photos.

The paragraph discusses the testing of Flux 1.1 Pro with more complex and challenging prompts compared to other top image generators like Ideogram version 2 and Mid Journey version 6.1. The presenter sets the image size to 1024x1024 and uses a series of difficult prompts to test the generators. The first prompt is for a woman doing a Warrior 1 yoga pose at home, which none of the generators handle perfectly, but Google's image gen 3 and Dolly 3 perform slightly better. The presenter notes that Flux 1.1 Pro and Flux 1 Pro produce very similar results, and Mid Journey fails to generate the correct hand and finger details. The second prompt is for a man giving a TED talk with a neon sign and specific lighting details. Flux 1.1 Pro understands the prompt well, generating the TEDEX logo and text correctly, although with some minor flaws. The comparison shows that Flux 1.1 Pro and Flux 1 Pro are quite similar, with minor improvements in 1.1 Pro.

This paragraph compares Flux 1.1 Pro with other image generators using prompts that aim to produce realistic, low-quality images. The presenter tests the generators with prompts like 'lowquality Snapchat photo of a Teenage man taking a mirror selfie' and 'a teenage woman holding a handwritten note that says verify me 8 2022 2024 lowquality selfie photo'. The results show that Flux 1.1 Pro and Flux 1 Pro produce images that are too polished and blur the background too much for the desired low-quality effect. Ideogram performs better in these tests, producing images that look more like real, low-quality selfies. Mid Journey struggles with text generation and overall realism. The paragraph also includes a sponsorship mention for Turbo Type, a tool for creating custom keyboard shortcuts to save time on repetitive tasks.

The paragraph focuses on testing Flux 1.1 Pro's ability to generate detailed images with prompts like 'two hands making a heart symbol' and 'a woman showing her palms and soles of feet'. The results for the heart symbol prompt are good, with Flux 1.1 Pro producing a realistic image, although the background is too blurry for an amateur photo. When testing the prompt for a woman showing her palms and soles of feet, Flux 1.1 Pro fails to meet the prompt's requirements, producing an image that does not show the requested details. The comparison with other generators shows that Ideogram comes closest to the prompt, while Mid Journey misses key elements entirely. The paragraph also notes that Flux 1.1 Pro's images tend to be slightly brighter and more saturated compared to other generators.

This paragraph evaluates Flux 1.1 Pro's ability to generate images in different styles, starting with a prompt for a watercolor painting of a whale in the sky. Flux 1.1 Pro successfully creates a watercolor-like image that follows the prompt closely. When compared to other generators, Flux 1.1 Pro is preferred for its watercolor effect and accurate depiction of the whale. The paragraph then tests Flux 1.1 Pro's ability to generate anime-style images with the prompt 'an anime girl in the city at night'. While the result is good, the text on the signs in the background is not realistic, which is a common issue with all current image generators. Comparing the generators, Ideogram produces a more anime-style background, while Flux 1.1 Pro and Flux 1 Pro generate images that look like anime illustrations or posters. Mid Journey's result is considered the least anime-like.

The paragraph tests Flux 1.1 Pro's ability to generate images of uncommon animals and different styles. Using the prompt 'a kodo dragon in the forest', Flux 1.1 Pro fails to generate a realistic kodo dragon, which is consistent with its known weakness in generating uncommon animals. Mid Journey outperforms the other generators in this case, producing the only recognizable kodo dragon, although with some inaccuracies. The presenter also experiments with the prompt 'komodo dragon' followed by a random date in JPEG format, but the result is still not a realistic kodo dragon. The paragraph concludes with a summary of the test and comparison of the four best image generators, noting that Flux 1.1 Pro does not show a significant improvement over Flux 1 Pro and shares similarities in performance with its predecessor.

This paragraph discusses the performance of Flux 1.1 Pro based on a third-party evaluation and its updates. Flux 1.1 Pro has the highest Elo score and win rate of 69% according to the image generator leaderboard by Artificial Analysis. It is revealed that Flux 1.1 Pro, previously known as the mysterious 'blueberry' model, outperformed all other models. The paragraph highlights that Flux 1.1 Pro generates images six times faster than Flux 1 Pro while improving image quality, prompt adherence, and diversity. Flux 1 Pro has also been updated to generate images twice as fast as before. Flux 1.1 Pro will soon be able to generate ultra-high-resolution images up to 2K without compromising prompt following, and its pricing is cheaper than Flux 1 Pro at 4 cents per image. The paragraph concludes by encouraging viewers to try the CR2 or JPEG prompts for generating realistic, amateur-looking photos and to share their experiences and prompting techniques.