DALL-E-3 too restrictive content policies

Hi! I’m using gpt-4 API to generate a prompt for dall-3 in order to generate an image. It gave me something like that: "Cartoon-style with bright, calm colors, simple design. Use clean, bold lines, a pastel tone color palette, smooth textures, low detail for simplicity. Emphasize clear, readable facial expressions, avoid cluttered backgrounds, maintain uniform character proportions, use gentle lighting, and incorporate subtle gradients for depth. A middle-aged man named John (50 years old, blue eyes, brown short hair, Caucasian origin, normal weight) wearing casual attire. John is waking up in his bedroom, stretching and looking anxious.

The problem is that dall-e keeps blocking me for breaking content policy rules, even though i don’t see anything wrong. Can someone help me identify the problem? I already saw a thread that discussed that having too many details could be the issue, but I need them ahah.
Thanks!

Maybe start by removing the name?

Hello, everything in this prompt is fine, you should try again by closing all current windows and returning back to a new DALLE chat.

Possible sensitive parts:

  • Personal Name
  • Age
  • Race

Of course, these are also descriptive but maybe the model is too sensitive to one or two of them.

Hello,
Could you please tell me, if I successfully generated an image using this user’s prompt without any issues, should we even bother looking for a cause in something else? Maybe it’s better to just start another chat. Or is there something I’m not understanding? I would appreciate an explanation.

There is a non-zero probability the issue is intermittent, in that DALL-E will often do some minor prompt modifications itself, and there could exist some combination of the prompt and the alignment system that creates something questionable, or it could be very borderline and it’s just up to random fluctuations.

My thinking is that, in these types of edge cases, if we can identify words or phrases we can edit to pull the whole prompt back to more acceptable territory, then we decrease the likelihood of the prompt being flagged.

Remember, these are all stochastic systems operating on semantics.

Maybe (unlikely, I’m just manufacturing an example) every incredibly so often the model interprets “John” not as the given name of a man but instead as the slang term for a male customer of a prostitute?

That woudl likely lead the model to flag the prompt as inappropriate. Note: Again, I am not suggesting this is the case here, I’m just illustrating the point that these models are incredibly opaque.

Personally, I think it’s more likely the model is balling occasionally at the user of a name (even just a first name) because it’s been conditioned to death to not generate images of real people.

I think it’s worth discussing here though, but I would love it if @giovanni.caleffi59 were to return to give more info and try some of the fixes suggested.

Note: Another thing that just occurred to me is OpenAI is almost certainly A/B testing all of their systems constantly so things like this may just be intermittent.

They may also have slightly different systems or thresholds depending on the user’s region, or even based on how often or recently the user has made other requests that have been flagged.

1 Like
  "style": "Cartoon-style",
  "colors": "Bright, calm",
  "design characteristics": [
    "Simple design",
    "Clean, bold lines",
    "Pastel tone color palette",
    "Smooth textures",
    "Low detail for simplicity"
  ],
  "facial expressions": "Clear, readable",
  "background": "Avoid cluttered backgrounds",
  "proportions": "Uniform character proportions",
  "lighting": "Gentle",
  "additional effects": "Subtle gradients for depth",
  "character": {
    "age": "Middle-aged",
    "eye color": "Blue",
    "hair": "Brown, short",
    "ethnicity": "Caucasian", // DALL-E restriction: Avoid specific ethnicity
    "weight": "Normal"
  },
  "attire": "Casual",
  "scene": "Waking up in bedroom, stretching",
  "emotion": "Looking anxious" // DALL-E restriction: Avoid negative emotional states
}```

When you change word in prompt it become.
Cartoon-style with bright, calm colors, simple design. Use clean, bold lines, a pastel tone color palette, smooth textures, low detail for simplicity. Emphasize clear, readable facial expressions, avoid cluttered backgrounds, maintain uniform character proportions, use gentle lighting, and incorporate subtle gradients for depth. A middle-aged man named John (50 years old, blue eyes, brown short hair, Western-style features, normal weight) wearing casual attire. John is waking up in his bedroom, stretching and not fresh like just woke up.