Mengchen West Wind comes from the Au Fei Temple
Quantum Position | 公众号 QbitAI
The latest AI Wensheng graph model Flux1.1 has been swiped overnight.
With just a simple trick, you can remove the "AI flavor" in the picture, and achieve a photorealistic effect regardless of people or landscapes.
The reaction of netizens in the comment area be like: I can't tell the difference, I really can't tell the difference.
This trick is also very simple to use, just imitate the file naming format of a DSLR camera in the prompt.
For example, "CR2" is the original image file format used by Canon cameras, enter "IMG" + random number + ". CR2", plus the specified content, you can get a realistic image.
Later, there were also feedback from netizens who tried it, and they could get good results by changing to the "ARW" of Sony cameras, "NEF" of Nikon cameras, and even Apple's "HEIC" format.
So much so that some people began to wonder if it could be a real photo from a random training data spit out by the model, right?
However, if you zoom in and look at some specific details, it is still easy to see that it is indeed AI-generated, such as the text on the license plate number is garbled.
So is the Flux 1.1 model inherently powerful, and how much of a role does this play in this?
A senior retoucher posted a comparison, adding IMG_1018.CR2 on the left, and not adding it on the right, he thinks the difference is huge.
Our actual results can also show that this technique can significantly improve the realism of the picture.
If you want to try this Flux 1.1 model for free, you can go to the together.ai platform and sign up to get $5 credits.
Randomly generate a selfie of a tourist on the Great Wall, which looks like that at first glance, but if you look closely at the texture of the character's skin, the mountains and plants in the background, there is still an AI smell.
换成“IMG_0314.cr2: selfie on The Great Wall”，是不是立马就不一样了？
Codenamed Blueberry, the latest SOTA Wensheng diagram model
With the official release of FLUX1.1, the mystery of the two unclaimed "blueberry" models that previously topped the Wensheng diagram model rankings has also been revealed, and it is it.
On the Artificial Analysis image arena, FLUX1.1 [pro] is codenamed "blueberry", surpassing all other models and obtaining the highest overall Elo score.
In contrast, FLUX1.1 [pro] is also cheaper and faster, surpassing Midjourney, SD3, Ideogram, etc.
In terms of generation speed, FLUX1.1 [pro] is 6 times faster than its predecessor, FLUX.1 [pro], while maintaining improved image quality, command response, and variety.
By the way, now FLUX.1 [pro] has also been updated, 2 times faster than before, and FLUX1.1 [pro] is 3 times faster than FLUX.1 [pro] currently available.
In addition, officials say that fast high-resolution generation is coming soon, and FLUX1.1 [pro] can natively support the kind that can generate 2k images without sacrificing any command response.
FLUX1.1 [pro]将通过Together.ai 、Replicate 、fal.ai、Freepik等在线平台提供。
At the same time, the official BFL API has also been launched, which can be integrated into their own applications by other developers, and the API pricing is:
- FLUX.1 [dev]: 2.5 cents per image (about RMB 0.18)
- FLUX.1 [pro]: 5 cents per image (about RMB 0.35)
- FLUX1.1 [pro]: 4 cents per image (about RMB 0.28)
Stable Diffusion原班人马打造
FLUX1.1 [pro]的背后是Stable Diffusion原团队,成员包括Robin Rombach、Andreas Blattmann、Dominik Lorenz等。
△Robin Rombach
In fact, Stable Diffusion started as an academic research project.
由Björn Ommer教授领衔,Robin Rombach、Andreas Blattmann、Dominik Lorenz等几位慕尼黑大学机器视觉与学习研究小组成员,以及Runway的研究人员Patrick Esser完成。
Seven months after the research paper was published, Stability AI stepped in to provide computing resources to further develop text-to-image generation models. In 2022, the authors of the above paper joined Stability AI.
团队一同创建了Stable Diffusion XL、Stable Video Diffusion等。
ICML 2024最佳论文之一、Stable Diffusion 3技术论文的Rectified Flow Transformers，还有SDXL-Turbo使用的对抗扩散蒸馏方法（Adversarial Diffusion Distillation），也是这帮子人在研究。
In March this year, it was revealed that these core research team members resigned en masse.
They then formed a new team called Black Forest Labs, based in Germany.
今年8月初刚刚宣布成立,并发布了其初代文生图模型FLUX.1,FLUX.1有三个变体:FLUX.1 [pro]、FLUX.1 [dev]和 FLUX.1 [fast],平衡了性能和可访问性。
目前黑森林实验室已完成种子轮融资,共融资3100万美元,Andreessen Horowitz领投,Brendan Iribe、Michael Ovitz、Garry Tan、Timo Aila和Vladlen Koltun等跟投。
They are also said to have received follow-on investments from General Catalyst and MätchVC.
Black Forest Labs and Musk have also collaborated to bring their image generation models into xAI's Grok assistant.
Next, the team revealed that it will launch a SOTA-level text-to-video generation model.
They are said to be raising $100 million at a $1 billion valuation, a significant increase from the previous $150 million valuation.
From Pika 1.5 to Meta Movie Gen, the video generation track exploded in the second half of this year, and the addition of the Black Forest Lab may bring a different spark.
Flux 1.1 demo
https://api.together.ai/playground/image/black-forest-labs/FLUX.1.1-pro
Reference Links:
[1]https://x.com/fofrAI/status/1841854401717403944
[2]https://blackforestlabs.ai/announcing-flux-1-1-pro-and-the-bfl-api/
[3]https://techcrunch.com/2024/10/03/black-forest-labs-the-startup-behind-groks-image-generator-releases-an-api/