
Google has quietly launched its latest text-to-image generator, Imagen 3.
Imagen 3 has already made waves by outperforming top competitors like DALL-E 3, Midjourney v6, and Stable Diffusion 3 in human performance evaluations.
Imagen 3’s strength lies in its ability to create high-quality, realistic images that closely match detailed text descriptions. However, it faces challenges with tasks requiring numerical reasoning, understanding scale, and depicting actions.
Users can try Imagen 3 through ImageFX and Vertex AI, experiencing its enhanced capabilities in generating various styles, from photorealistic images to abstract artworks, with less effort in prompt engineering.
Google’s decision to release Imagen 3 without much fanfare might be a strategic move, considering the excitement surrounding Black Forest Lab’s open-source Flux-1.
Safety and ethics have been central to Imagen 3’s development, with extensive filtering and data labeling processes to minimize harmful content in the training dataset.
Google has also incorporated its SynthID watermarking technology to embed imperceptible markers in generated images, aiding in identifying AI-created content.
This approach contrasts with models like Flux, which allow users to fine-tune the AI with custom datasets and do not include built-in watermarking.
While this openness has sparked interest in the AI community, it also raises concerns about potential misuse and identifying AI-generated content.