Summary of Key Points
Recently, there have been two major developments in the field of AI-generated images:
1. Imagine Image 2.0, promoted by Elon Musk, has gained popularity for its "precise layer editing" and "emoji pack" features, ranking second in the world in image generation and editing. However, it has faced criticism from users due to its strict review process.
2. OpenAI is quietly testing a new version of GPT-Image 2 (codenamed mona-lisa-1), which offers significantly improved image realism and comes with an OpenAI-specific watermark. Although the knowledge base is updated only up to May 2025, it is suspected to be a simplified or open-source model.
Detailed Analysis
1. Imagine Image 2.0: The Dark Horse Rising with Practical Features
The success of this model lies in its precise layer editing capability, which allows users to separate an image into multiple layers (e.g., the clothes, face, and background) without needing to know how to use Photoshop. This makes it easy for non-experts to make edits:
- Editing emoji packs: You can replace elements in a classic "distracted boyfriend" image with a cup of milk tea or change the boyfriend's gaze to look at a game.
- Exporting transparent backgrounds: Simply select the main subject (e.g., a cat) and download a background-free image, which can be used for creating emoji packs or posters.
- Combining multiple images: Upload up to 5 images, and the model will automatically merge them into one (e.g., combining a landscape, a person, and text).
Elon Musk has been highly promoting Imagine Image 2.0 on X.
2. The Dual Edge of Imagine: Practicality vs. Strict Review
While the advantages are clear—Imagine offers a more user-friendly interface compared to GPT-Image 2 (which requires writing detailed instructions)—the strict review process can be frustrating. For example, uploading an image of Spider-Man may result in a copyright violation error, and even slightly off-topic images are rejected.
3. OpenAI's New Model: More Realistic, but with Limited Knowledge
Users have noticed that OpenAI is testing mona-lisa-1, which features:
- Improved realism: Images generated by this model look more authentic, resembling those taken with a smartphone (with motion blur and uneven lighting).
- Watermark identification: The images come with an SynthID watermark to verify they were created by OpenAI.
However, there is a downside: the knowledge base only covers up to May 25, 2025. For instance, if you ask the model to draw the timeline of Anthropic Inc., it may produce incorrect information from later dates. Some speculate that this could be a simplified version of GPT-Image 2 or an open-source model being prepared by OpenAI.
4. The Evolution of AI Image Generation
The focus in the competition among AI image generation models has shifted from producing the most visually stunning images to creating more practical and user-friendly tools:
- Interactive editing: Tools like Imagine's layer editing make it easy for non-experts to modify images.
- Realism enhancement: New models aim to mimic the look of smartphone snapshots, reducing the "artificial" feel associated with AI-generated images.
- Practicality: Models should meet users' real needs, such as creating emoji packs and editing meeting room visuals.
Users are no longer just interested in seeing interesting images; they want tools that are actually useful.
5. Future Expectations for AI Image Generation
People are hoping that AI image generation will improve in the following areas:
- Consistency:生成的 images should maintain consistent character appearance and style (e.g., a character in a comic series should remain recognizable across different images).
- Clear text: Text should be clearly visible even when generating images with large amounts of text.
- Creative aesthetics: Models should be capable of producing stunning artistic styles (e.g., Chinese or cyberpunk themes).
- Flexible editing: Users should have more control over combining elements from different images or having the AI automatically fill in missing parts.
AI image generation is moving beyond just showing off technical capabilities and is becoming increasingly practical, meeting the everyday needs of ordinary users.