Hi guys, im trying to avoid getting references to any brand like now, the Pepsi brand is coming up very often. Is there any way for this brand to enter a Black List and avoid references to this brand, no matter how small it may be ?
Thanks.
1 Like
Any idea about this problem ? Any ideas or tips to improve this problem?
I’ve posed this question to one of the Dall-E team and negative prompts are not a thing, at least currently, no mention of a timeline where that would be included either. I can only advise you to perhaps include in the prompt not to include brand names.
1 Like
I saw something about there being no negative prompts although I think they are one of the most important things. About the brands, it happens to me that with the simple fact of talking about, for example, a COLA drink, or even a soft drink, I make reference to the PEPSI brand in its colors and logo, maybe not 100% direct but at a glance you can see the brand. An important issue if we want to develop applications where the customer demands to avoid any mention of their competition. I feel you should emphasize this.
ChatGPT (or the API) works different.
It’s not about negative prompts or seeds or temperature, it’s all about semantics, context and taxonomy.
The system is “very bad” at things you don’t want, so you have to focus on things you do want.
Do not say “the sun is not visible” or “there is no sun”, but say “it is dark” (reverse psychology).
3 Likes
I have had good results by telling it not to suggest, recommend, or endorse other brands. I agree that this is something OpenAI needs to focus on more. When AI assistants are representing a brand, they cannot be discussing any other brands at all unless they’re some sort of sponsor.
I Understand thanks. Lately I test using this technic and work fine. I test Focusing on colors I want to see and not what I don’t want. Also with this I can control the abuse of some colors when I want to print and this work.
1 Like
Try adding word like “IMPORTANT: < something here >”, or capitalizing “DO NOT …”.
Things like this mostly work for me, let me know if it works for you, curious about whether it’s generally good strategy or not.
1 Like
That is absolute the best approach.
Do not say “I don’t want cold colors” but do state “the image depicts only warm and vibrant colors”.
Don’t say “do not use Pepsi” but say “use Coca-Cola and Fritz-Kola only”.
Another worthy tip, indeed.
Do not say “the car and bike are both visible, but the bike is not so important” but state;
The car is the most important thing in the image, it is shown with sharp focus on the foreground, where the bike is unsharp and with a very soft focus on the background.
E.g. ChatGPT / Dall-E is stupid, very stupid - remember that! Speak to him like “a thing” that can’t do anything, like a toddler.
Don’t tell him what you don’t want, say what you want.
Double negative are even more difficult, like “the car is not unimportant”… say “the car is very important!”.
2 Likes
Certainly what you mention is what works. I have tried generating images following these tips and I have better results.
Thank you so much!!
Close your eyes, visualize the image you want to see.
Tell GPT what you see.
Describe the tone, the room, the ambience, the composition, the visible materials, the shadows, the lightning, the atmosphere, the “look and feel” of the image, the style, the era is was created in, etc…
It sounds simple, but good prompting is really difficult (well, for me at least).
1 Like
It is true, one is not used to dreaming and detailing what is being visualized. But I think it’s just practice. We hope that the future will be less complex or easier to parameterize. Anyway, I’m seeing, analizing and testing other tips. I hope find more good ideas.
1 Like
Well put. Negatives for me just make a mess. Point it in the right direction and you can get great results.
The word ‘only’ seems very powerful. In system messages ‘Your only role is to [thing]’ is a miracle worker for me.