Recently, several "TED speakers" went viral online, but it was revealed that none of them are real people. These hyper-realistic AI-generated images fooled millions of viewers and even AI detection software.
The images were created by Leo Kadieff, a former member of the Stable Diffusion team, using the latest Flux RealismLora technology. This LoRA model greatly improves the realism of AI-generated images compared to previous methods.
Key points:
- The images are raw outputs without any upscaling or post-processing
- They use a 22MB RealismLora file and ComfyUI workflow
- The technique simplifies prompts - a simple "RAW hyperrealistic photo, UHD, 8k" is sufficient
- Comparison images show a dramatic improvement in realism with the LoRA model
- Even Midjourney struggled to replicate the level of photorealism
The viral post revealing these were AI-generated reached over 5 million views in just hours. Many viewers expressed shock at how indistinguishable from real photos the images are.
This demonstrates the rapid progress in AI image generation, raising questions about the future ability to distinguish AI from reality. The technology is becoming accessible to average users, not just AI experts.