Getting Chatgpt to understand?

[Lucan Unlordly]

I've been using Chatgpt for image watercolour colorization and the first results are ok..........but not perfect.....and I cannot get it to understand the requested adjustments?

For example: The first draft is very good but the facial features need tweaking. The subjects teeth weren't showing as they were in the original upload.
The second attempt and the subject now looks like a different person and is holding a bag whereas they had gloves on in the original and the first image.

This debacle continues until it asks me: If you want your exact photo to stay identical (same hand, glove, lighting, everything) but just to be colorized and stylized in a watercolor effect, I can do that properly by editing your uploaded image directly rather than regenerating it.

YES PLEASE!

The bag remains, the facial features now make the subject look like Benny Hill and I tell it it's blxxdy useless

 

[BusterBloodvessel]

I can't help you, but I do feel your pain! I've really started to get into using ChatGPT in the last month or so, in fact I use it far more than Google now. Mainly for summarising information, getting insights, rewriting emails etc..... but yes, images it's like it suddenly has the brain of a 4 year old and the memory of a goldfish!! I've experienced exactly what you say - a first edit that's 90% there, then in the next revision it fixes the final 10% you requested but then completely makes a balls up of everything else that was otherwise perfect!

As I say, can't help you.... but at least you know you're not alone

 

[fisicx]

ChatGPT can’t create anything new. It doesn’t learn. All it can do is assemble an image based on the data it has slurped from millions of existing galleries.

Which means unless someone has already painted a picture that exactly matches parts of what you want you are unlikely to get the result you desire.

 

[Ozzy]

There are many different 'AI' tools out there, and ChatGPT isn't the recommended one for image work. It's primarily focused on written content, like writing content or researching papers.
For image work, I'd recommend using different tools such as Midjourney or Stable Diffusion, although apparently, Google is now getting very good at image and video.

 

[DontAsk]

What did you expect?

I saw a 3d model supposedly generated from a photo by AI. The problem was the subject had his hands behind his back and they turned into an amorphous pink blob.

 

[AmazonGeek]

The problem is the prompting. So get Chat to generate the 'meta prompt'. Tell it what you want to do and ask it for the prompt. The more detail you give it the better. Some other tips...
- if you are starting with an image (e.g. a product that you want in a lifestyle setting), don't describe the product itself. Only reference the environment you want it to appear in
- Use Chat GPT 5 'thinking' (use the selector to switch between the different modes)
- Say "max quality' in the prompt. Chat has 3 settings and it will pick the one it thinks is most relevant for the situation but you can force max quality
- say what the size is if this is relevant.
- once it generates an image it is very difficult to get it to change a minor part of it. More often than not it will mess up something else in the process. So if it is 90% there, use another programme like Canva or photoshop to tweak it.

 

[fisicx]

The alternative is to ask a graphic designer, artist, photographer or otherwise to do it for you. Keeping them employed and not supporting copyright infringing tech behemoths in the US.

 

[Ozzy]

Keeping them employed and not supporting copyright infringing tech

That's a nice ideology, but the reality is that the landscape is changing, and AI tools aren't going to disappear, so learning how to use them effectively is the wisest choice.
Not all these tools are effective, and like driving a car, some people can master a car through a rally course with perfection and others crash trying to reverse out of a parking space.

 

[Lucan Unlordly]

The alternative is to ask a graphic designer, artist, photographer or otherwise to do it for you. Keeping them employed and not supporting copyright infringing tech behemoths in the US.

Can I ask you to handwrite your reply and stick it in the post to me?

The original images are mine and I've been buying and using Corel Draw for over 20 years. Unfortunately I'm as thick as 2 short planks and can only do so much with it!

 

[mikewilliam474]

Improve your prompts

 

[Lucan Unlordly]

Improve your prompts

I'm grindin' frfr but I ain't even close yet, Innit!

 

[AmazonGeek]

use chat gpt to write the prompts. Tell it what you want to do and it will generate a 'chat-friendly' prompt for you.

 

[Ozzy]

use chat gpt to write the prompts. Tell it what you want to do and it will generate a 'chat-friendly' prompt for you.

I have done this a few times when needed to do repetitive work. Spent ages getting the output right, and then asked the GPT to write a single prompt to get me where I was in one prompt instead of 50

 

[GLAbusiness]

It kept giving me answers which needed PowerShell or cmd. I told it I preferred windows GUI solution where possible and it adjusted its responses

 

[rebecca778]

I've been using Chatgpt for image watercolour colorization and the first results are ok..........but not perfect.....and I cannot get it to understand the requested adjustments?

For example: The first draft is very good but the facial features need tweaking. The subjects teeth weren't showing as they were in the original upload.
The second attempt and the subject now looks like a different person and is holding a bag whereas they had gloves on in the original and the first image.

This debacle continues until it asks me: If you want your exact photo to stay identical (same hand, glove, lighting, everything) but just to be colorized and stylized in a watercolor effect, I can do that properly by editing your uploaded image directly rather than regenerating it.

YES PLEASE!

The bag remains, the facial features now make the subject look like Benny Hill and I tell it it's blxxdy useless

I have tried using many AI tools for image editing, and I have encountered the same problem as you mentioned. Personally, I think there are two reasons. One is that the AI can't understand what you said, and the other is that the AI cannot recognize certain details of some images. So you either have to try hard, or accept a mediocre result, or use Photoshop to retouch the image yourself. I hope my reply was helpful to you.

 

[JEREMY HAWKE]

It will get there remember it has to do this in every language on the globe then fathom out what you really mean ,Then try to deliver what you really want
We maybe asking for to much at this embryonic stage

 

[FreddyG]

We maybe asking for to much at this embryonic stage

No. You are asking the wrong AI. The clue is in the title - LLM. Large Language Model. It specialises in language, not pictures.

LLMs are great for politics, history, economics, idiomatic translations, stuff like that. Though the Google thing thought that a Great Dane in German was a magnificent Danish person!

(I just wish that our magnificent Danish person would not crap in the hallway! Our German sheep farmer doesn't do it, but then she sheds hair all over the place.)