Introduction
When you use AI to generate images like "a boy playing games on a phone", you might get strange results — like the boy holding the phone backward, just so the game screen is visible. Even though AI can create sci-fi blockbusters, it struggles with simple scenes like this. This article will break down why this happens and give you practical tips to fix it.
Why Does This Happen? The "Prompt Pollution" Problem
The real issue lies in your prompt. AI has a "keyword association trap" (or "prompt pollution"): when you mention a word, even a negated one, AI automatically fills in related stereotypes.
For example:
- If you say "generate a person wearing a nurse cap", AI will not only add the cap but also a nurse uniform and a hospital background.
- Even if you say "generate a person not wearing a nurse cap", AI still puts them in a hospital.
- "Not wearing a safety helmet" → AI places the person at a construction site.
- "Not wearing a Christmas hat" → AI creates a Christmas scene.
- "Not wearing a chef hat" → AI shows a kitchen scene.
The word "game" in your prompt becomes a "pollutant". AI prioritizes showing the game scene, even making the boy hold the phone backward to display the screen.
How to Fix It: Avoid the "Pollutant" Word
The solution is simple: play a word game. Instead of "a boy playing games on the subway", rephrase it to "a boy sitting on the subway, holding the phone horizontally with both hands". By removing the "game" pollutant, the problem is solved.
Advanced Tip: Reverse the Trap
You can even use this trap to your advantage. For example, when generating "a boy editing photos on a computer":
- If you say "a boy looking at the front of the monitor", AI flips the screen so you (not the boy) can see it.
- Instead, say "a boy looking at the back of the monitor". This "ungrammatical" sentence for humans makes AI generate the correct image.
Another example: Instead of "a man with eight-pack abs", use "a man with broad shoulders and narrow waist". This avoids the "abs" pollutant and still creates a fit look.
The "Visual Compensation" Bug
Another issue is the "visual compensation" bug. AI lacks a perfect world model and can't understand abstract actions like "facing away" or "looking at". It compensates by drawing concrete visual elements:
- "Facing away" → AI draws the back, backpack, and helmet.
- "Looking at" → AI draws the eyes, visor, and front face.
When both "facing away" and "looking at" are in the prompt, AI compromises by showing a side view.
How to Fix It: Describe Visual Elements Directly
To generate an astronaut facing away from the camera and looking at Earth, avoid ambiguous words. Instead of "an astronaut standing on the moon, facing away from the camera, looking at Earth", use "an astronaut standing on the moon, facing away from the camera, with a planet Earth floating in the distant space". This describes the visual elements (the astronaut's back and Earth) instead of abstract actions.
Advanced: Control the Camera
You can use visual compensation to control the camera:
- To zoom in on an astronaut floating in space, describe details like "worn metal buckles, flashing indicator lights on the chest control panel, clear nylon woven texture on the arms". AI will focus on these details, effectively zooming in.
- For a close-up, describe the astronaut's face: "clear eyes through the visor, complex iris texture in the pupils, fine hairs on both sides of the nose".
- To change the shooting angle, describe elements from that angle. For a top-down view, say "the round top of the helmet, the shoulder contour visible from above, the top connector of the oxygen cylinder".
Final Tips
When generating images, avoid the mistake of first building the entire space and then placing the camera. Instead, describe what you can see in the scene. This will greatly reduce your chance of "drawing failures".
常见问题
Why does AI add things I explicitly told it NOT to include?
This is the "negation paradox" in prompt engineering. AI models don't process negation the way humans do — when you say "no hat", the model activates the concept of "hat" first, then tries to suppress it. The activation is stronger than the suppression, so hat-related elements (hairstyles, head accessories, sometimes even the hat itself) leak through. The fix: never use negation in image prompts. Instead of "a person without glasses", say "a person with clear, unobstructed eyes". Instead of "no background clutter", say "minimalist white background, empty space". This is the single most impactful habit change for better AI image generation.
Does this advice apply to all AI image generators (Midjourney, DALL-E, Stable Diffusion)?
Yes — the prompt pollution and visual compensation behaviors are universal across all major image generation models. They stem from how these models are trained (text-image pairs), not from any specific model's architecture. Midjourney is slightly more resilient to pollution because of its stronger aesthetic training, but it still exhibits the same patterns. DALL-E 3 is more literal and benefits most from the "describe visual elements directly" technique. Stable Diffusion is the most sensitive to prompt pollution and benefits most from the "avoid pollutant words" technique. The camera control trick (describing details to force zoom) works across all three.
How do I know if my prompt has a "pollutant" word?
Ask yourself: "If I remove this word, do I still convey the same visual?" If yes, the word is likely a pollutant. "A boy playing games" → "A boy holding a phone horizontally with both hands" — the visual is the same, but the pollutant "game" is gone. The test: generate both versions and compare. If the second version is more accurate, you found a pollutant. Common pollutants to watch for: action verbs (playing, eating, working), abstract roles (doctor, teacher, gamer), and emotional states (happy, sad, angry). Replace them with visual descriptions: what the person is physically doing, what they're wearing, what their expression looks like, what's in the environment.
Can I use these techniques for video generation too?
Yes — the same principles apply to AI video generation (Sora, Seedance, Runway, etc.) with one additional consideration: motion. For video prompts, you need to describe both the visual elements (using the techniques in this article) AND the motion. Example: instead of "a person walking happily", use "a person walking forward at a steady pace, arms swinging gently, hair bouncing with each step, slight smile on the face, bright sunlight casting moving shadows". The motion description replaces the abstract "happily" without introducing a pollutant word. The camera control trick (describing details to control framing) works for video too, but you'll also want to specify camera movement explicitly: "slow push-in, handheld style, slight vertical bounce".