Google 研究揭示 Nano Banana 图像生成的五类使用模式

A new Google research paper shows how the company can identify what users are doing with Nano Banana without violating user privacy. They subsequently discovered five different ways image generation is used.

The Problem And Solution

People are uploading sensitive images of themselves and their families to Nano Banana so the concern about privacy is real.

Yet, AI image generation is wildly popular and understanding how people use Nano Banana is crucial to identifying how it’s applied and what can be done to improve it for the different and emerging ways it’s being used. So the challenge was to identify a way for understanding how Nano Banana was being used while also respecting user privacy.

The research paper explains:

“Beyond proactive self-reports in traditional media and on social networks, there is limited data on how and why many people are using LLM image generation.

With this growth in usage and emergent use cases, we sought to provide some descriptive analysis of Nano Banana use, to help understand the ways in which users might find tools like Nano Banana helpful for social connections, educational materials, scientific advancement, commerce, and more.”

The researchers proposed that they can understand how people use AI image generation by analyzing patterns in their text prompts and conversations, without needing to examine the uploaded or generated images themselves, while protecting user privacy.

The researchers used an existing system called Urania, an existing framework Google already uses for gaining insights from LLM chat interactions.

Analysis & Findings

The researchers analyzed 2.1 million Gemini Nano Banana conversations and grouped them into five categories:

1. Social & Viral
2. Brand & Marketing
3. Product Visualization & E-Commerce
4. Architectural & Spatial Design
5. Educational & Technical Information

They also compared how people used the technology across those five categories, examining:

1. Artistic styles
2. Image editing features
3. Image composition features
4. Interactions/session features
5. Multi-select Features

1. Image Editing Features

This activity includes removing objects and people from an image, modifying a background, adding new elements, and modifying images by resizing or restoring them.

“Image Editing included explicit requests for modifications to alter, clean up, or change specific elements within an image that has already been created (or a reference image uploaded by a user). “

2. Artistic Styles

Artistic styles is another interesting feature where researchers discovered that users were asking for specific art mediums like watercolors, pencil sketches, photography style, or images like oil paintings. Other users vibed their images by describing moods or styles like minimalist, film noir, or vintage. Some of the more creative users requested conceptual mashups that paired clashing styles or moods.

The researchers explained: 

“Artistic Style included explicit user requests about the creative direction and visual aesthetic, from the artistic medium used to

specific stylistic or cultural influences.

…Absurdity or Humor:
Requests that intentionally incorporate elements of nonsense, satire,

surreal comedy, or humorous exaggeration.

Imitate Art, Person, or Character:
Explicit requests to emulate the distinct style of a specific artist, public figure, or fictional character.”

3. Image Composition

Within the five category of uses they also noticed that users composing the images with lighting, perspective, constraints (telling it what not to do), and requests for people and how many.

“Composition refers to requesting specifics about the structural arrangement of the scene, including placement of subjects, perspective, lighting, etc”

4. Interactions/Session Features

“Interactions features look at the conversation as a whole and how people iterate on and edit generated images.”

5. Multi-Select Features

These classify conversations across additional dimensions like prompts that include text and typography requirements, types of image transformation, work context, and stages of content creation. Multi-select Features provide additional information about what people are trying to accomplish and how they use image generation.

The multi-select features includes:

• Full Specification
  Whether users precisely specify a style and wording.
• Precise Content Only
  The user provides the text content but not the style.
• Stylistic/Generic Only
  The user prompts contain the vibe of what they want but without specifics.
• N/A
  The prompt doesn’t mention text or fonts.

What The Researchers Discovered

The researchers found that Nano Banana was being used for more than trivial viral social media uses. People were also using it for commercial activities, educational illustrations, technical instructions, and product visualizations.

The most surprising finding is that memes and humorous images were the least prevalent features in the social and viral categories. They found that the most prevalent feature was the inclusion of people (nearly 80%). Requests involving image composition and artistic styles were also common, while prompting the LLM to start over was about twice as prevalent as asking for minor changes.

The researchers noted:

“Perhaps surprisingly, meme style or format and absurdity / humor are the least prevalent.”

Lastly, the researchers discovered that much of the brand and marketing activity involved original, fictional, or unrecognized brands rather than established companies. The researchers suggested that some users may be developing branding for new or speculative businesses.

They pointed to these findings related to the marketing usage of image generation:

Marketing Images Emphasized Precision

Marketing usage emphasized specific visual requirements. The most prevalent features in marketing conversations involved controlling how images were composed, including the placement of elements, people, lighting, and instructions about what should not appear.

Marketing Had The Highest Prevalence Of Detailed Text Requests

Compared with the other four categories, marketing had the highest prevalence of requests specifying exact text and detailed typography instructions.

This suggests that marketers are using image generation to create promotional images with specific text and design requirements rather than simply requesting attractive images.

Small Businesses And Individuals More Prevalent Than Big Brands

The Brand & Marketing category had the highest prevalence of conversations identified as involving small businesses, individuals, or side hustles.

Marketing Uses Beyond Generating Images From Scratch

The researchers identified activities involving refining logos, changing text and colors, creating promotional materials, and developing consistent visual brand identities, including assets for online engagement and traditional promotional uses.

Takeaways

The researchers successfully used the Urania framework to conduct a “privacy-preserving descriptive analysis” to identify what users were doing with Nano Banana. They discovered that AI image generation is enabling a variety of business, marketing, and creative activities, including some that tasks that previously required professional experience that is now within reach of more people.

They also acknowledged that it remains unclear whether AI is enabling entirely new activities or simply making existing ones less expensive. They expressed that they anticipate additional uses to emerge as people continue experimenting with AI image generation.

Featured Image by Shutterstock/Khosro