I have been reading/seeing/trying numerous examples since months and came to a conclusion.
Images made using Chatgpt can clearly be marked as unreal (not realistic) though using best prompt to make it look like real.
On the contrary trying to use an existing photo of myself, trying to play with prompts, my face/photo person looks real but rest of things appear computer-generated (Photoshoped)
What am I doing wrong and need to improve ?
Would love to see some examples, where a pompt makes realistic image, which becomes difficult for human eye to differentiate.
An incredible variety of imagery can come out; make yours a photo and describe how it was taken, and then don’t hesitate to spend more time than the image takes to create writing what you want:
This is a photograph taken with a modern DSLR camera with a 35mm portrait lens, and subtle fill flash was used to fill shadows a bit from the natural daytime sunshine. The image contents are a prominent young man and a woman in the front row behind the stage barricade in a small outdoor concert audience among a background of others rock concert attendees, where the two of them have attended the concert with ridiculous outfits. The man is dressed as a southern aristocrat colonel in a white plantation suit and is wearing a white top hat about two meters tall (to the chagrin of some of those immediately behind). The woman is dressed in a beaded sequined 1920s flapper gown in black and gold shiny accents, and has a flapper’s headdress with one of the largest ostrich feathers in the cap that you’ve ever seen, also flapping back and forth to bother the other concert attendees. They are aware of having their picture taken and are making garish poses for the camera. The other hundreds of attendees are frazzled sweaty classic alt-rock fans with looks from grunge to goth to summerwear for the sunshine, and certainly didn’t expect the scene made by these two in the front. There are no unnatural items in the photo merely as text placeholders which over-describe the theme or situation: this is a picture taken at the alternative rock concert scene event by an event photographer.
The above I typed up without going back, and pasted here before I sent - let’s see if this is clearly made by AI…
Send to “instant” ChatGPT - you don’t need the intercepting “chat” AI thinking too hard about the high-quality language you’ve already written yourself.
If you are me: Receive a persistent error I’ve been getting in ChatGPT, when going between past messages, when going between GPT items and starting a chat. And be thankful that you “pasted it here” so you don’t have data loss from your work typing away.
Recover: press Enter on the written chat URL in the browser bar again for a complete reload, now with a query string ?mweb_fallback=1 appended, whatever that means, and go back to your chat where things have been chugging away for you, and evaluate your new image - as unreal as the silly people’s outfits you described:
When you are uploading an image, the AI image model will be doing its best to only replace parts of the image described, almost identically placed. If you are the type that wants to send your face into OpenAI’s products, then you’ll still likely want to prompt something along the lines, “infer my face and appearance alone, and place me in an all-original image and brand new pose, not reciting back directly from my provided image” and more language along those lines. That way you don’t get a “bad photo edit” with merely surroundings changed.
It would help if you could share your prompt here, so we could better help you to improve.
But you already got really great advice here.
Since I have my profile picture visible here, I actually did a test with ChatGPT couple of weeks ago, editing my profile picture, just to see if it is believable or if it’s noticeable that it is AI generated. My prompt was simple: “Edit so I have bangs and brown eyes”.
I’m showing these as an example, since it is what your question is about, making your edited appearance look believable.
I am unable to post multiple pictures in one post and I can’t post a link to the conversation.
You may not get the best result in 1 image go. But you should be able to produce what you need if you keep at it, giving the chat GPT direction and commenting on what’s wrong.
Definitely describe your issue and talk about it before asking it to draw something.
My prompts were too basic to mention, hardly 2 to 3 lines
With keywords realistic, surreal, etc. keywords to make those photos look real life. But failed.
I have noticed, there is heavy use of dark red/brown color in all photos background or most used color at back. (or maybe my observation)
I can find similar images on internet, were people are using ChatGPT to generate those photos/image
Trying prompts from those topics, makes my images/photos different compared to what is posted by user, not sure why though ?
Also, the most used colors like red/cherry/brown is common in all photos, generated at my end.
Would it be possible for you to post examples of the images you’ve prompted? It would make it easier for other users to understand what feels different in your outputs and help suggest prompt changes that might get you closer to the visual result you want.
This is a great place to ask, share and learn from each other and this community has many knowledgeable users.
This how I work with prompts. I actually don’t usually start with very complex prompts. My own workflow is more iterative.
I often begin with a fairly simple prompt that describes the main idea, mood, subject or style I want. Then I look carefully at the output and adjust based on what feels wrong, missing or interesting. This goes for images/videos.
So for me, prompting is less about writing the perfect prompt from the start and more about working with the model through the result:
Start with the core idea.
Generate.
Look at what the model understood well.
Notice what it misunderstood or overdid.
Adjust only the parts that need changing.
Repeat until the image/video gets closer.
For example, if the image looks too artificial, I wouldn’t only write “make it realistic.” I’d look at why it feels artificial: lighting, skin/texture, background, proportions, colors, too many effects, etc. Then I’d correct that specific thing.
For regular work, I use the same idea: start simple, see what the model gives, then guide it step by step instead of trying to solve everything in one prompt. But as said, this is how I work and this workflow fits me.
