DALL-E Adds Unwanted Elements Despite Clear Instructions.

[limber night]

 I'm encountering a persistent issue with DALL-E where it consistently adds unwanted elements to the images it generates, for example regarding T-shirt graphic designs gpt. Despite having clear instructions to create standalone designs without any T-shirt in the image, DALL-E repeatedly places the design on a T-shirt. This happens even when the prompt doesn't mention a T-shirt.

More puzzlingly, when attempting to correct this, DALL-E acknowledges the instruction but still proceeds to include a T-shirt in the subsequent attempts. This issue persists regardless of the level of detail in the instructions. I'm curious to know if others have faced similar challenges and if there are any known workarounds. Any insights or suggestions to address this would be greatly appreciated.

[hollow fog]

In the current version of DALL·E 3, negative prompting ("Don't include x") only confuses the model for bringing it up in the first place.

DALL·E can't see what it makes, and it just doesn't support "negation" in general, at least of yet. So current best practice is to keep descriptions focused solely on what you do want to see.

[limber night]

Thanks for sharing this insight about DALL-E 3 and negative prompting. It makes sense now why the model was including elements I wanted to avoid. The idea that mentioning something, even in a negative context, can actually trigger its inclusion is really interesting, and admittedly, a bit challenging.

I had a similar issue with T-shirt designs. After removing any mention of 'T-shirt' from the instructions, I still faced the same problem. It seems that even having 'T-shirt' in the project name might have been enough to trigger its inclusion in the design. This can be tricky for users who might unintentionally use language that prompts DALL-E to include these unwanted elements.

Interestingly, I found that when I explicitly instructed GPT to fix the image and specifically avoid mentioning the keyword that triggered the unwanted element, it helped. This approach of adjusting the GPT instructions for DALL-E seems effective, but it's not immediately obvious. Most users might continue trying to directly remove elements, not realizing this subtlety in how DALL-E interprets

[upper musk]

Dall-e and all diffusion models work by associating words to images. The best way to get an image is to try to figure out how it would be visually described.

(note this is a simplification)

[limber night]

Got it, that makes sense. I'm just worried once the app store starts growing and popularity increases, the average user isn’t going to know how to prompt correctly for specific apps. I am sure most would say something along the lines of create a t-shirt design or mug design before their prompt. It would be nice if there were an area to give detailed instructions or guidelines on how to promote to get the desired outcome on that app.

[limber night]

Hey, I've got a question about giving instructions to a GPT. Mine seems to have a thing for making lists, which I get, but I'm really aiming for more of a natural, flowing conversation style. It started doing the list thing even before I gave it any specific instructions. Do you think the following instruction would help steer it away from lists, or might it just end up doing the opposite? Just trying to figure out the best way to guide it. Any advice would be super appreciated!

'I avoid using numbered lists or bullet points, favoring a narrative style that mirrors a warm, personal conversation.'"

[hollow fog]

GPT is more capable of understanding negative prompting than DALL·E is, especially GPT-4, so I think you're on the right track with your instruction. A couple bits of feedback I'd offer, and I'm sorry I'm gonna just bullet them 😁

• Try using custom instructions, if you're not. This way, ChatGPT will keep it in mind every time it's generating a new response.
• Phrase the custom instruction in the second-person, instead of the first-person, so that ChatGPT understands you're telling it what you want it to do, instead of just sharing a fact about yourself, which it might be assuming now based on the phrasing.
• This one is probably more optional, but might help a bit - add a little more import to your request with something like: "It's important not to respond using numbered lists or bullet points. It's important to favor a warm and personal narrative and conversational style."

[limber night]

Thanks a lot for your helpful advice! 😊 I've actually been using custom instructions, so it's great to know I'm on the right track there. Your suggestion about phrasing the instruction in the second person makes a lot of sense. It’ll definitely help clarify that I'm directing the gpt, rather than sharing something first person .

And I really like the idea of adding more emphasis to the instruction. That sounds like a smart way to ensure gpt understands the importance of avoiding lists and maintaining a warm, conversational style. Your insights are incredibly valuable, and I'm grateful for the guidance! I always appreciate someone taking the time to help out.

[hollow fog]

You're very welcome, I'm glad you found it helpful! Depending on your current use case too, one last bit of advice I can offer is to explore making a custom GPT, if you haven't yet. This is kind of like the "next level" of custom instructions, and your custom instructions are actually disabled when using custom GPTs. But, you have 8000 characters to work with, instead of 1500+1500, and extra functionality like the ability to upload knowledge files for the model's permanent access, and more advanced "Actions" functions too.

[limber night]

Thanks for the great advice on custom gpt! I have created about 12 custom models so far. It's been quite an adventure fine-tuning each one with that generous 8000 character limit for the system prompts. It really lets you get into the nitty-gritty of how they operate and respond.

I'm particularly excited about integrating unique knowledge files and experimenting with the advanced Actions functions. There's so much potential in tailoring these models for specific tasks. Just hoping they don't cap the number we can create, right?
Anyway, cheers for the tip! It's always cool to meet someone else who's into customizing these gpt. 🙂

[hollow fog]

Absolutely! Bonus tip: many users have reported that the "load more" button of GPTs doesn't work right when getting about to the # you've already made, which can limit your ability to navigate back to them easily, so it might be worth manually bookmarking them yourself or otherwise saving the URLs!

[limber night]

I noticed that and was wondering about that i’ve been having to find a conversation with that specific gpt on the menu. That's a smart move, bookmarking them or saving the URLs. Thanks for the heads-up on that! It's these little tips that can really make a difference. 😊👍