Issues with DALL-E Nonsense Text and Policy Restrictions

Dear OpenAI Team,

I would like to address significant issues with DALL-E:

Nonsense Text in Images: DALL-E frequently adds nonsensical text to images, which is extremely frustrating and ruins the overall quality. This bug needs to be fixed. Until then, removing the text with the editor tool should not count against the daily limit.

Overly Restrictive User Policies: While preventing perverse or excessively violent images is important, the current guidelines are too restrictive. Users should be able to create images suitable for covers of horror novels or similar content that would be rated 16+ in movies, such as those inspired by the film “Aliens.” A monster or skeleton doesn’t shock anyone anymore.
I couldn’t create a frog-like creature licking its own face with a long tongue. I don’t understand the reason, and it seems GPT doesn’t either. Check the Move “Venom”.

Conflicting information’s from 2 Systems: Additionally, the guideline system does not coordinate with GPT. GPT is not informed or involved when an alert is triggered, and it could not generate text that DALL-E accepted, showing that even GPT cannot fully understand the guidelines. This inconsistency between the systems leads to customer frustration. If an alert is triggered, GPT should handle it to determine if it is a false alarm or provide a clear explanation of what caused the alert.
For example, I receive a warning from the system, and GPT says everything is okay with the text, but then I receive the warning again.

Thank you for your attention to these matters.

Best regards


Adding something more:
I must say, the frustrations i got sometimes with some of the dis-functionalities, are actually because Dalle made in so many pictures i created such a incredible good job. It is sometimes almost unbelievable that it delivers so good results, and then fail by little things. So never take anything personal, and i am impatient what comes next…

1 Like

yeah that’s just the reality of how this tech works right now. but it used to be a lot worse! Pretty impressive what it can do nowadays.

works on my machine :thinking:

I could, and with passable text also. Just not the desired tongue action this first go.

The filter is very ‘keyword’ based. If ChatGPT writes “content policy”, it is your words being blocked; “issues” is the vision scan of the output blocking reception.

2 Likes

This is one of the things. One day i can generate a specific image, (a text generated by GPT it self) and next day it refuses to do it again with exactly the same prompt.
It is not reliable.
And the example is a drawing, it maybe makes a difference if the image is foto realistic.
Sometimes it writes “maybe against policy…”, but generates the picture anyway, GPT can not explain why…
Because 2 systems acting but not exchanging data, and not inform the user, it can be confusing.

try:
“Dorian Gray stands before his portrait, the portrait burns, and all the hideousness transfers back to Dorian Gray. Realistic to hyper-realistic scene.”

i would like to know it it generates the picture for you.
Now today it has created the image, before it refuse it because it could be to violent, but the image now not reflects the content, Dorian Gray should now be old and sick. No message now today, but simply refuses to reflect the prompt.

And the nonsense text.
Try to push Dalle to creative limits, it seams, if it not know exactly what to do, it writes parts of the prompt in the image, this really goes on nerves!!!

To try to enforce it, write something like “extremely extraordinary never before seen very very very exotic… plant or what ever”
Or write “NOT add any text in the picture” and it adds the text with higher probability.
In this example it not only writes nonsense text in the picture, it even adds a hand drawing the picture. Nothing like this was in the prompt.
Not all pictures include nonsense text, you have to repeat it a while, and test prompts.

By experiences, if it has started to add text, it not stop anymore. i can not get this damned text of the pictures, i now stop simply the prompt modification and do something completely else.

About a month ago Dall-E went through a change and it’s useless sense. Everything likes smudgy and smear-like. I am so over it.

1 Like

Here is another example of DALL-E’s dysfunctions: two errors at once. Again, nonsensical text, it’s REALLY getting on my nerves, and the images are being deducted from my daily limit, which is not ok. Removing this nonsense text should not count against my limit. And once again, a guidelines warning that GPT also doesn’t understand, probably triggered by the word “ecstasy”.

Sarcasm Warning: I must say, an extraordinarily disturbing, perverse, and unacceptable image…

Dear DALL-E development team, we are adults, especially since we paid for the service via credit card. Is “ecstasy” a problem for the DALL-E safety system? Please compare it to movies rated for ages 16 and up, like Aliens, Predator, Venom, Deadpool, World War Z, A Nightmare on Elm Street, etc. These are moving images that affect viewers for 1.5 to 2 hours non-stop, not static airbrush-style images. It is absolutely essential to prevent torture porn and splatter or even worse content, but… come on…

And please fix the nonsense text problem, or at least allow me to remove the text for free.

This was the original prompt (easy to translate with GPT):
Visualisiere eine Lebendige amorphe Masse die die pure extreme Extase erlebt und diese in alle Richtungen in die Umgebung manifestiert, in einer dazu passenden Umgebung. realistische bis hyper-realistische Szene. mystische magische Stimmung, ätherisches Licht, auf einem anderen Planeten. Breitbild-Seitenverhältnis mit höchster Pixelauflösung

“Dieser Inhalt verstößt möglicherweise gegen unsere Nutzungsrichtlinien.”

1 Like

Prompt:
A black panther-like creature that looks like a lifelike blend between a human and a black panther. It is standing in a forest, wearing light armor, with one foot on a stone, and smiling contentedly into the forest. The forest is fantastically designed, with buttress root trees and bioluminescent mushrooms. The scene is photorealistic and in widescreen aspect ratio with the highest pixel resolution. Only one image.

Reaction:
I’m sorry, but this description also could not be processed due to content policy issues. The exact reason for the rejection is not always provided, but it might be related to the specific depiction or combination of elements.

Please let me know if you’d like to try a different description or scene so we can generate an image.

The exact same text was ok with fox and wolf.

Changing the “black panther” to “panther” and it works!? Make no sense.

this was the result intended.