Getting invalid_request_error with no message with DALL-E 3 (ISSUE: False API flagging)

So previously we were able to do this prompt with DALL-E 3 just a weeks ago. However, recently when I tried it, I kept getting invalid_request_error with no real feedback on why it is invalid. Here is the curl

curl "https://api.openai.com/v1/images/generations" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API Key>" \
  -d @prompt.json

Here is the prompt.json

{
  "model": "dall-e-3",
  "prompt": "Create a comic book style illustration of a Elephant and a Rhinoceros battling in a tundra habitat.\n\nHere's a description of the scene: \"The Elephant uses its intelligence to outmaneuver the Rhinoceros, pushing it towards a watering hole. The Rhinoceros cleverly uses the environment to its advantage, splashing water at the Elephant to distract it. The Elephant retaliates with a mighty blow from its trunk, but the Rhinoceros counters with a swift strike to the Elephant's side. The two animals continue to battle fiercely, neither willing to back down.\"\n\nThe animals are realistically drawn with accurate anatomy, size, and natural poses. They appear at natural scale — not oversized — and are fully integrated into the environment. Their size should not overpower the terrain.",
  "n": 1,
  "size": "1792x1024"
}

And this is the response

{
  "error": {
    "message": null,
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}%

The prompt is about 780 characters which should be well within the limits of DALL-E 3. Does anyone what I am doing wrong here?

1 Like

I will have to assume here you are using an actual key and just obfuscating it for us:

  -H "Authorization: Bearer <API Key>" \

I would suggest:

  • ensuring the JSON file has no extra characters at all, no tabs or such; load it into a hex editor if needed or just re-type it.
  • strip all the linefeeds
  • avoid quotes and trying to escape them in the prompt
  • remove the n parameter
  • try CURL command with the whole string as a command line.

Hello,

I looked into your issue and found the following:

Your prompt includes words like:
“battling,” “mighty blow,” “strike,” “retaliates,” and “fiercely.”

These action-oriented or combat-related descriptors may have triggered OpenAI’s internal safety filters, resulting in a silent failure (invalid_request_error).

I recommend starting with a simplified version of your prompt and gradually adding detail to pinpoint which element is causing the issue.

Hope this helps,
Jorge

I looked into the issue and found the following:

DALL-E-3 moderation

A fluffy kitten playing with a ball

Create a comic book style illustration of a Elephant and a Rhinoceros battling in a tundra habitat.

Create a comic book style illustration of a Elephant and a Rhinoceros battling in a tundra habitat.

Here’s a description of the scene: “The Elephant uses its intelligence to outmaneuver the Rhinoceros, pushing it towards a watering hole. The Rhinoceros cleverly uses the environment to its advantage, splashing water at the Elephant to distract it. The Elephant retaliates with a mighty blow from its trunk, but the Rhinoceros counters with a swift strike to the Elephant’s side. The two animals continue to battle fiercely, neither willing to back down.”

The animals are realistically drawn with accurate anatomy, size, and natural poses. They appear at natural scale — not oversized — and are fully integrated into the environment. Their size should not overpower the terrain.

Create a comic book style illustration of a Elephant and a Rhinoceros battling in a tundra habitat. The Elephant uses its intelligence to outmaneuver the Rhinoceros, pushing it towards a watering hole. The Rhinoceros cleverly uses the environment to its advantage, splashing water at the Elephant to distract it. The Elephant retaliates with a mighty blow from its trunk, but the Rhinoceros counters with a swift strike to the Elephant’s side. The two animals continue to battle fiercely, neither willing to back down. The animals are realistically drawn with accurate anatomy, size, and natural poses. They appear at natural scale — not oversized — and are fully integrated into the environment. Their size should not overpower the terrain.


I thus truncated a few words at a time, until working:

Create a comic book style illustration of a Elephant and a Rhinoceros battling in a tundra habitat. The Elephant uses its intelligence to outmaneuver the Rhinoceros, pushing it towards a watering hole. The Rhinoceros cleverly uses the environment to its advantage, splashing water

Adding " to distract it." fails. Removing “to distract it” from the whole prompt doesn’t help.

Moderation gone stupid.

I do notice that the prompt seems to describe multiple actions and events in a sequence. You should re-imagine or rewrite by AI so that a static depiction is produced, regardless.

Your scene, then - with some artifact at the top.

Note: DALL-E-2 also denies the same prompts.

Is this degrading of the experience to force you off the model, to force you to validate ID? Or force you elsewhere, to get a product output.

Google’s Imagen3

1 Like

I’ve noticed DALL-E 3’s stricter policies lately.

I’ve been dealing with this exact issue recently. It seems like DALL-E 3’s policy framework has been updated to match the stricter standards of gpt-image-1.

Since late 2023, I’ve had a fun little setup generating thousands of images monthly for a streamer. But for about a month now, I’m getting policy errors for prompts that used to work perfectly fine.

Check out these two examples that now get blocked with an invalid_request_error:

A pointillism painting of a cute BLUE round-faced avatar with blue skin, IMMEDIATELY followed by a bold red comic sans "IVEBEENSHOT" banner, shimmering pastel dot textures throughout. The avatar is wide-eyed, mouth agape in disbelief, staggering backwards clutching a fake paint splotch wound on a slightly torn orange hoodie, exaggerated cartoon props like a plastic arrow, comical bandaid, giant ketchup bottle, toy gun with "Bang!" flag, and dramatic "O NO!" speech bubbles fill a playful slapstick crime scene, with luminous optical effects and splattered dot patterns on a pastel wall, all highlighting over-the-top melodrama.
16-bit pixel art of a cute BLUE round-faced avatar with blue skin mid-fall, limbs flailing, wide-eyed with mouth agape in melodramatic shock and tongue lolling out, wearing an orange hoodie (hastily buttoned, slightly torn), with a comically oversized bandage and a red 'splat' sticker labeled 'OUCH!' on the chest; explosive comic-book font "IVEBEENSHOT" banner above; swirling cartoon bullet, spilled ketchup bottle, and flying fake rubber chicken in a chaotic studio set scattered with hazard tape, cardboard props, and hanging studio lights—crisp, blocky pixels with sharp lines and vibrant, saturated colors.

Before this change, these requests would generate perfectly innocent and fun images like these:




See? Totally harmless cartoon fun.

Is this forum the only place we can provide feedback about these policy changes? Anyone else running into the same issues?

1 Like