Google’s nano banana just killed Photoshop... let’s run it

Key Concepts

• Gemini Flash 2.5 (Nano Banana): Google's new image model.
• Image consistency: Maintaining the original character's appearance during image alterations.
• AI-generated assets: Images or sprites created using AI models.
• SynthID: Invisible watermark on Google-generated AI content.
• Real-world understanding: The model's ability to generate realistic images based on real-world data.
• Uncanny valley: The unsettling feeling caused by AI-generated images that are almost, but not quite, human-like.

1. Introduction of Gemini Flash 2.5 (Nano Banana)

• Google released Gemini Flash 2.5 image model, nicknamed "Nano Banana."
• The model is positioned as a potential disruptor to Adobe Photoshop.
• It allows users to make image alterations using prompts, maintaining consistency with the original image.
• It is suggested that this model can lead to significant productivity gains for animators and graphic designers.

2. Capabilities and Features

• Performance: Nano Banana is described as an exceptional image model, topping the LM Marina leaderboard.
• Cost: It's affordable, costing 3.9 cents per image via the API.
• Character Consistency: The model can blend images or make minor changes without altering the original character. Example: Blending 13 different images together while maintaining character consistency.
• Applications:
  • Creating LinkedIn headshots without needing a suit.
  • Implementing Mandela effects (e.g., removing the cornucopia from the Fruit of the Loom logo).
  • Generating character positions for animations.
  • Creating entire sprites with a single prompt.
• Real-World Understanding: The model can generate realistic photos based on Google Maps data.
• Sketching: Users can sketch step-by-step and the model will generate images based on the sketches. Example: Mapping out AWS infrastructure using sketches.
• Text Handling: The model can turn images into ads or memes.

3. Limitations and Drawbacks

• Text Errors: The model sometimes adds extra characters to words.
• Prompt Following: It may not always follow prompts exactly, doing its own thing or ignoring the prompt.
• Uncanny Valley: Character consistency is overhyped, as the model still produces images with an "AI uncanny valley" look.
• Censorship: The model is highly censored, preventing the generation of NSFW content.

4. AI Asset Disclosure and SynthID

• Users must disclose any AI assets used when publishing on platforms like Steam.
• Google-generated content has an invisible watermark called SynthID.

5. Learning AI with Brilliant

• Brilliant is the sponsor of the video.
• Brilliant's "How AI Works" course teaches users how to build a language model from scratch.
• The course allows experimentation with feature vectors to edit facial expressions and images.
• Brilliant offers a hands-on approach to learning.
• A QR code and a URL (brilliant.org/fireship) are provided for viewers to try Brilliant for free and get 20% off their premium annual subscription.

6. Conclusion

• Gemini Flash 2.5 (Nano Banana) is a powerful and affordable image model with impressive capabilities, particularly in character consistency and real-world understanding.
• However, it has limitations, including text errors, prompt following issues, the uncanny valley effect, and censorship.
• Learning how AI models work under the hood can help users get the most out of these tools.