Google's AI just CRUSHED image editing

Key Concepts

• Nano Banana/Gemini 2.5 Flash: A multimodal AI model with image generation and editing capabilities.
• Multimodal Model: An AI model that can understand and process various types of data, including natural language, audio, video, and images.
• Deep Fakes: Synthetic media where a person in an existing image or video is replaced with someone else's likeness.
• Style Preservation: The ability of an AI model to maintain the original style and characteristics of an input image during editing.
• Virtual Try-On/Clothes Swapping: The ability to change the clothes of a person in an image with another outfit.
• Photo Restoration: The process of repairing and enhancing old, damaged, or low-quality photos.
• Color Correction: Adjusting the colors in an image to make them more accurate or appealing.
• LM Marina & Artificial Analysis Leaderboards: Platforms where users can blind test and rank different AI models.
• AI Studio: Google's platform for using and experimenting with AI models.
• API (Application Programming Interface): A set of rules and specifications that software programs can follow to communicate with each other.

Nano Banana/Gemini 2.5 Flash: The Best AI Image Editor

Introduction

The video introduces Nano Banana, revealed to be Gemini 2.5 Flash, as a superior AI image editor. The video aims to demonstrate its capabilities, compare it with other leading image editors (GPT-4o and Quen Image Edit), and show how to use it.

What is Nano Banana?

• Nano Banana is Gemini 2.5 Flash with image generation and editing capabilities.
• Gemini 2.5 is a multimodal model that understands natural language, audio, video, and images, and can also create images.
• According to the LM Marina leaderboard, Gemini 2.5 Flash significantly outperforms other AI models, leading by almost 200 ELO points.

Deep Fake and Reference Image Accuracy

• Nano Banana excels at capturing details from reference images, making it suitable for creating deep fakes.
• Example: When given an image of Will Smith and Emma Watson in an art gallery and prompted to create a midshot photo with specific clothing, Nano Banana accurately generated the image, preserving the likeness of the actors and the details of the art gallery background.
• GPT-4o refused to generate the image due to censorship issues.
• Quen Image Edit generated an image but failed to replicate the background accurately.

Pose Understanding and Image Generation

• Nano Banana can generate images based on sketches of poses.
• Example: When given a sketch of fighting poses and prompted to generate an image of Will Smith and Emma Watson fighting in those poses, Nano Banana created an accurate image.
• Quen Image Edit failed to understand the prompt.
• GPT-4o understood the prompt but produced an image with distorted faces and less accurate poses.

Complex Design and Model Sheet Generation

• Nano Banana can generate model sheets with front, back, and side views of complex characters.
• Example: When given an image of the Destiny Gundam and prompted to create a model sheet, Nano Banana generated a detailed and consistent model sheet with accurate details across all three views.
• Quen Image Edit failed to generate some details, had flaws in the design, and produced inconsistent views.
• GPT-4o also failed to generate consistent views and accurate details.

Outfit Swapping and Style Preservation

• Nano Banana can accurately swap outfits between people in an image while preserving the original style.
• Example: When given an image of people and a pig and prompted to swap the outfits of the humans and turn the pig's head into a chicken's head, Nano Banana successfully performed all the edits while maintaining the original style.
• GPT-4o only turned the pig into a chicken but did not swap the outfits.
• Quen Image Edit turned the pig into a chicken but changed the style of the image and failed to swap the outfits.

Clothes Swapping/Virtual Try-On

• Nano Banana excels at clothes swapping, allowing for virtual try-ons.
• Example: When given an image of a man and an anime character and prompted to change the man's outfit to the anime character's outfit, Nano Banana accurately swapped the outfits while preserving the man's face and pose.
• GPT-4o applied the outfit but changed the person's face.
• Quen Image Edit failed to apply the outfit.

Race Swapping

• Nano Banana can swap the races of people in an image while preserving other details.
• Example: When given an image of an Indian man and a white woman and prompted to swap their races, Nano Banana accurately swapped their races while preserving everything else in the image.
• GPT-4o did not swap the races correctly.
• Quen Image Edit failed to understand the prompt.

Photo Restoration

• Nano Banana can restore and colorize old, blurry, and damaged photos.
• Example: When given an old black and white photo and prompted to restore and colorize it, Nano Banana restored the photo, preserving the details of the people's faces and correcting the damages.
• GPT-4o subtly changed the faces and added extra fingers.
• Quen Image Edit performed well, making it a close tie with Nano Banana in this case.

Color Correction

• Nano Banana can restore the colors of photos with messed up brightness, contrast, saturation, and white balance.
• Example: When given a photo with intentionally distorted colors and prompted to restore the colors, Nano Banana accurately restored the colors while preserving the details of the original photo.
• Quen Image Edit also did a good job, but Nano Banana's colors were slightly more accurate.
• GPT-4o changed too many details of the image.

Complex Scene Generation

• Nano Banana can generate complex scenes based on detailed prompts.
• Example: When given an image of a character and prompted to create a scene with a figure of the character, a box with the character's image, and a computer showing the Blender modeling process, Nano Banana accurately generated the scene.
• Quen Image Edit generated a similar scene but with less realism.
• GPT-4o messed up the face and outfit of the character and the details of the screenshot.

Manga Page Colorization and Unexpected Behavior

• Nano Banana can colorize black and white manga pages.
• Unexpected Behavior: In one example, Nano Banana unexpectedly changed the panels of the manga page to a completely different image, even though it usually excels at preserving details. The generated image seemed to reference an "evil form" of the character, suggesting a deeper understanding of the source material.
• Quen Image Edit colorized the page and kept the text the same but did not know the character.
• GPT-4o colorized the page but subtly messed up the details.

Limitations and Failures

• Lens Simulation: Nano Banana failed to accurately simulate different camera lenses.
  • Example: When prompted to retake a fisheye photo with a 100mm lens, it failed to remove the fisheye effect. GPT-4o performed better in this case.
• Expression Generation: While capable, Nano Banana sometimes struggled to accurately portray specific emotions.
  • Example: When prompted to create a grid of photos with different expressions, some expressions were not accurately portrayed. GPT-4o performed better in accurately portraying the expressions.
• Object Removal: Nano Banana did not always completely remove objects from a scene.
  • Example: When prompted to remove tourists from a photo, it left a few people in the background.
• Complex Transformations: Nano Banana struggled with complex transformations like turning girls into cats and boys into dogs.
  • Example: It failed to accurately transform all the girls into cats and boys into dogs. GPT-4o performed relatively better.
• Style Transfer: Nano Banana struggled with certain style transfers, such as turning a photo into anime style.
  • Example: It failed to turn a realistic photo into anime style. GPT-4o and Quen Image Edit performed better in this case.
• Homework Assistance: Nano Banana could not accurately label organelles in a cell diagram.
  • Example: It failed to label the organelles correctly and had typos.

Successes

• Text Editing: Nano Banana can accurately edit text in images while preserving the original font and style.
  • Example: It successfully changed "naturally" to "with poison" in an image.
• Text Translation: Nano Banana struggled with translating English text into Chinese.
  • Example: The Chinese translation was incorrect. GPT-4o and Quen Image Edit performed better in this case.
• Zooming and Detail Enhancement: Nano Banana can zoom into a photo and enhance details, but sometimes changes the pose.
  • Example: When prompted to zoom in on a kingfisher, it changed the pose of the bird. Quen Image Edit performed better in preserving the original pose.

Where to Use Nano Banana/Gemini 2.5 Flash

• Google's AI Studio: Available for free with a limited quota.
• LM Arena: Available for blind testing and direct chat.
• API: Available for developers to integrate into their apps.

Benchmarks and Leaderboards

• Google's benchmarks show that Gemini 2.5 Flash outperforms GPT-4o, Flux Context, and Quen Image Edit in most categories.
• LM Marina and Artificial Analysis leaderboards rank Gemini 2.5 Flash as the top image editor, leading by a significant margin.

Conclusion

Nano Banana/Gemini 2.5 Flash is a powerful AI image editor that excels at preserving details, understanding prompts, and creating deep fakes. While it has some limitations, it generally outperforms other leading image editors in various tasks. The video concludes by expressing excitement about the potential of future versions of Gemini and encourages viewers to share their experiences with Nano Banana.