AI resolution:AI 해상도

[洛雨]

1. Visual details: Look for "unnatural" flaws

1. Logical errors in human/biological details

- Hand and finger deformities: AI often mishandles the number of fingers, joint bending, and nail shape (such as extra fingers, finger adhesion, and joint dislocation), which is one of the most typical vulnerabilities.

- Uncoordinated facial features: such as different eye sizes, misaligned pupil reflections, blurred lip textures, chaotic tooth arrangement, or stiff expressions (muscle movement that is not a real expression‎).

- Hair/texture faults: The transition between hair and scalp and animal hair is unnatural, or repeated and blurred textures appear (such as hair suddenly "disappearing" or color faults).

2. Object structure and perspective contradictions

- Disproportionate proportions of buildings/furniture: such as abnormal door and window sizes, chaotic number of stair steps, and incorrect perspective of furniture corners (the rule of near big and far small fails).
​
- Missing or misplaced objects: for example, a cup has no handle, a book lacks cover texture, a vehicle wheel is not in proportion to the body, or the direction of the object's shadow is inconsistent with the light source.
​
3. The "deliberate feeling" of natural scenes
​
- Abuse of repeated patterns: Textures such as grass, tiles, and leaves appear to repeat regularly (similar to pixel replication), rather than random distribution in the real world.
​
- Light and color discontinuity: The light and shadow at the edge of the object are abrupt (such as the shadow is suddenly cut off), or the color transition is unnatural (such as the sky gradient color block).

2. Technical tools: with the help of AI detection and reverse search

1. AI-generated image detection tool
​
- Dedicated detection platform: Some tools identify by analyzing image generation features, for example:
​
- Google DeepMind's Identify (under testing), Truepic (for news pictures), Hivemind (detection of diffusion model generation pictures).
​
- Domestic tools such as "Wenxin Yige" and "Zhipu AI" have also launched generation picture detection functions (pictures need to be uploaded).
​
- Principle: AI-generated images differ from real photos in pixel distribution and noise patterns, and detection tools analyze these underlying features (but have limited effect on advanced models).
​
2. Reverse image search
​
- Upload images to platforms such as Google Images, Baidu Images, and TinEye to see if there are the same or similar image sources. If the image has no network records and the content is not a rare scene, it may be AI-generated (but please note: newly generated images may not be included).
​
3. Metadata (EXIF) analysis
​
- Right-click the image to view "Properties" or "Details". If the metadata shows that the camera model, shooting time and other information are missing, or shows "AI generation tool" related software (such as DALL·E, Midjourney, Stable Diffusion), it may be AI-generated. However, some generation tools will delete metadata, so it is necessary to combine other methods to judge.

3. Logical reasoning: judging from content and scenes

1. Counter-common sense content or details

- The picture contains a combination of objects that do not exist in reality (such as "flying cars + ancient buildings"), or the details violate the laws of physics (such as water flowing to high places, objects floating without support).

- Text content errors: If AI-generated pictures contain text (such as road signs, newspapers), garbled characters, spelling errors, and semantic incomprehension (such as "It's sunny today").

2. The "perfection" and "blur" of the scene coexist

- AI may portray the subject very delicately, but the background details are blurred or simplified (such as the main character is clear, and the background trees and building outlines are blurred), forming "inconsistency between the primary and the secondary".

- Overly "idealized" scenes: For example, the colors in landscape photos are too saturated, the composition is too symmetrical, and there is a lack of randomness in the real world (such as the distribution of leaves and natural flaws in the movements of characters).

4. Advanced skills: Features of specific generative models

1. Diffusion model (such as Stable Diffusion) generation image

- There may be "artifacts": slight blurring or color overflow at the edge of the object (similar to the effect of low-resolution images after enlarging), or repeated patterns appear in complex textures (such as fabric patterns).
​
2. GAN model generation image

- Early GAN-generated faces may have "eyes lost" and "skin texture plastic", but the model has been greatly optimized in recent years, and it is necessary to combine other details for judgment.

5. Note: AI technology is advancing, and the difficulty of distinguishing is gradually increasing

- "Anti-detection" of advanced models: Some AI tools will deliberately add "realistic" noise (such as simulated camera noise, film grain), or optimize detail logic, which is difficult to distinguish by vision alone.
​
- Comprehensive judgment is more reliable: do not rely on a single method, and combine "visual vulnerability + detection tool + reverse search" for multiple verifications.
​
- Pay attention to the credibility of the source: If the picture comes from a non-authoritative platform, there is no clear shooting information, or the content involves rumors or exaggerated propaganda, you need to be vigilant.

Summary

The core flaw of AI-generated pictures lies in the "simulation defect of real-world logic". Whether it is a detail error, technical traces or content contrary to common sense, multi-dimensional observation + tool assistance is a more effective way to distinguish. With the development of technology, more professional detection methods may be needed in the future, but for ordinary users, catching "unnatural details" is still the most direct method.

thank you

[ILLEGAL]

조용히 해라 니네 나라가서 떠들어라

[洛雨]

이게 큰 소란으로 여겨지는지 묻고 싶습니다. 저는 단지 AI를 식별하는 기준과 기술을 설명하고 있을 뿐입니다.
()
😅

[Kyzand]

Xi Jinping, Son of a bitch!!!!

[洛雨]

Then I wish you to marry a "good" wife

[우선 너부터]

그러니까, 개화 일러스트는 Ai로 만들었을 가능성이 존재한다는 거군요.

이번 단편 스토리는 어떠셨나요?

[洛雨]

No, I just give developers the ability to identify AI works.

The story was boring and empty at the beginning, but then it had the same flavor as before, and the experience was very good afterwards