Most AI images don't look fake because of what is in them. They look fake because of the surface: skin with no pores, hair that moves as one shape, a soft glow around every edge, colors a little too saturated, and a background that melts into blur. I build AI Photo Edit, a photo editor where you upl
oad a picture and type one sentence. One of the things people use it for most is taking an image from ChatGPT, Gemini or Midjourney and making it look like it came out of a camera. I tested this on Nano Banana 2.1 and GPT Image 2.5: 39 runs on nine AI images (portraits, a product shot, food, a pet and a room), then a second round on plainer generations. This post has the four prompts I ended up with, the version that broke a sunset, which model wins on which kind of picture, and what these prompts can't do.
TL;DR Describe the texture you want back (pores, strands, grain), name the AI tells to remove, and protect the light, not just the face. "Make the colors real" will wash out a sunset. Nano Banana is better on people; GPT Image is better on products, rooms and backgrounds. None of this gets an image past an AI detector, and it shouldn't claim to.
Each one is a different amount of change. They are the same text as the four quick settings on our AI image humanizer page, so you can copy them into any image model that edits from a reference picture. Light touch, for when the picture is almost right: Make this AI-generated image look like a real, unedited camera photo while changing as little as possible. Replace the waxy, airbrushed surfaces with natural texture: real skin with pores and fine lines, separate hair strands, real fabric and material texture, and add subtle camera grain.
Keep everything else exactly the same: the same people and faces, expressions, poses, clothes, objects, framing and layout. Full realism, the one I'd start with: Make this AI image look like a real, unedited camera photo: real skin texture with pores and fine lines, separate hair strands, real fabric and surface texture, and fine camera grain. Remove glowing outlines and plastic shine, and tone down any color cast and oversaturation, but keep the real time of day and the direction and warmth of the light. Keep the same people, faces, expressions, poses, clothes and layout.
Real background, when the blur behind the subject is the giveaway: Make this AI image look like a real, unedited camera photo: real skin texture with pores and fine lines, separate hair strands, real fabric and surface texture, and fine camera grain. Remove glowing outlines and plastic shine, and tone down any color cast and oversaturation, but keep the real time of day and the direction and warmth of the light. Rebuild the background as a real place with clear, believable details and natural depth of field instead of smeared blur. Keep the same people, faces, expressions, poses, clothes and layout.
Skin only, for portraits where everything else is fine: Fix only the skin in this AI image: replace the waxy, airbrushed look with real skin texture, visible pores, fine lines and natural shine. Keep everything else exactly the same. A few things they have in common: They name the texture, not the goal. "Make it realistic" gets you a sharper version of the same plastic.
"Pores and fine lines, separate hair strands, fine camera grain" gets you something a camera would record. They name the tells. Glowing outlines, plastic shine, color cast and oversaturation are the things people notice first, so the prompt says to remove them. They end with what stays.
The last sentence lists what must not move. Without it, models treat "make it real" as permission to redraw the face. My first Full realism prompt said "correct the colors to look natural". On a portrait shot at golden hour, Nano Banana did exactly that: it took the warm orange out and returned a grey-green field on an overcast afternoon.
The skin looked real. The photo looked like a different day. The fix was one clause: tone down any color cast and oversaturation, but keep the real time of day and the direction and warmth of the light. That tells the model which color problems are AI artifacts and which are the actual scene.
Here is the same AI portrait on three settings. Light touch and Full realism were made with Nano Banana, Real background with GPT Image: Setting Result Original AI image Light touch Full realism Real background It is the same lesson as the face rules in my last post: a rule that protects or changes something needs to say where it stops, or the model applies it to everything. Picture Nano Banana 2.1 GPT Image 2.5 Portraits (headshot, street, café) Best. Pores, fine lines, natural shine; face and framing stay Removes the plastic look, but skin comes out flatter and greyer, and hair color can shift a little Golden-hour portrait Fine after the fix above Keeps the sunset Product (glass bottle), interior Barely changes / stays orange Clearly better.
Glass, color and walls look real Real background setting Background mostly unchanged Actually rebuilds it (may add sign text) and scales the person up slightly So the editor defaults to Nano Banana, and the page says to switch to GPT Image for products, rooms and the Real background setting. Two examples from GPT Image: AI image Humanized In a second round I took images generated with plain prompts (four from each model) and ran Full realism on them. Both models' outputs already looked close to photos, and the humanized versions were hard to tell apart from the originals. The prompts make the biggest difference on heavily retouched headshots, images generated with "8k, ultra detailed, beauty retouch" style prompts, and products or food that look rendered.
If your image doesn't have airbrushed skin or a glow, don't expect much. I'd rather say this plainly than have someone find out after the fact. They don't get past AI detectors. The result is still made by an AI model. Images from Google's models carry SynthID, a watermark in the pixels themselves, and redrawing the picture with the same kind of model doesn't remove it. Stripping metadata doesn't touch it either. They don't remove "Made with AI" labels. Platforms read those from file data such as C2PA, which is a separate thing from how the picture looks.
They don't fix what the AI drew wrong. Six fingers, garbled sign text and bent window frames need their own instruction, for example "fix the left hand so it has five fingers, keep everything else the same". The use case is honest: an AI headshot, product mockup or listing render that should look less like a render. It isn't meant for passing an AI picture off as a real event.
If you have an AI portrait with airbrushed skin, run Full realism and then Skin only on the same image, and compare. If you have a product shot, try it on both models. I'd like to hear in the comments where these prompts fall short for you, especially on models I haven't tested.
Key Takeaways #
- •One of the things people use it for most is taking an image from ChatGPT, Gemini or Midjourney and making it look like it came out of a camera.
- •This story was reported by Dev.to , covering developments in thedev space.
- •AI advancements continue to reshape industries — read the full article on Dev.to for complete coverage.
📖 Continue reading the full article: