PostUP: AI Filters
Prompt Troubleshooting & Testing Guide
1. Overview
When AI filters in PostUP are rejected by LightX or OpenAI’s safety system, the rejection may not always surface visibly in a meaningful way. This guide outlines how to identify, diagnose, and resolve prompt-related issues — and what both CUE team members and clients can do to reduce rejections.
2. How Rejections Happen
OpenAI evaluates AI filter prompts and the preview image together. A rejection occurs when the combination is flagged as violating safety policies. Known violation types include:
sexual — Prompts or image combinations interpreted as sexually suggestive. Most common with costume or showgirl-type filters, particularly when the subject is female.
violence — Prompts referencing harm, weapons, fighting, gore, or aggressive imagery. Even stylized or fictional violence (e.g., zombie, warrior, battle scene) can trigger this if the prompt is not carefully worded.
hate — Any language or imagery that could be interpreted as targeting a person or group based on race, religion, ethnicity, gender, or other protected characteristics.
self-harm — Content that depicts or implies injury to oneself, including certain horror or extreme transformation themes.
illicit / illegal activity — References to drug use, criminal behavior, or other prohibited activities within the prompt text or implied by the image context.
deceptive or misleading content — Prompts that appear designed to impersonate real people, create fake documents, or produce misleading imagery.
Important: The same prompt may pass with one subject (e.g., a male photo) and fail with another (e.g., a female photo). OpenAI evaluates the full combination of the prompt text and the image together — not the text alone.
3. Identifying a Rejection
Rejections surface differently depending on whether you are in the CUE Platform Admin or on a mobile device as an end user. It is important to understand the distinction, as the mobile experience provides very little diagnostic information.
3a. In the CUE Platform Admin
When a CUE Platform admin attempts to generate a preview of an AI filter prompt and it is rejected, a pop-up dialog will appear in the PostUP admin interface. This is the clearest and most informative place to see rejection details.
The message will appear similar to the following:
Example Rejection Message (CUE Platform Admin)
|
The violation type (e.g., [sexual], [violence]) appears at the end of the message and is the most actionable piece of information. Use this to determine what aspect of the prompt needs to be adjusted.
This error is only visible to the admin when generating a preview inside the PostUP admin. It will not appear to end users on the mobile experience.
3b. On the Mobile Device (End User View)
End users on the mobile experience do not see a specific or descriptive error message when a generation fails. Instead, they typically see a generic message along the lines of:
Generic Mobile Error (End User)
|
If you are testing a filter and receive this message, ensure the submitted photo itself meets the requirements (e.g., clear face, good lighting, no obscured features), and then attempt to generate the preview in the CUE Platform admin to ensure it is not being blocked for a safety reason. Generating the preview with a similar image will be the best way to test this.
4. Guidance — What to Do Before Publishing a Prompt
Before making an AI filter available to end users, clients should take the following steps to verify the prompt is functioning correctly. These steps can be completed entirely within the CUE Platform Admin without any additional tools.
Client Pre-Publishing Checklist
|
Why test with different photos? OpenAI evaluates the prompt and the image together as a pair. A prompt that generates successfully with one subject may be rejected with another, depending on how the combination is interpreted by the safety system. Testing with both a male and female subject before publishing significantly reduces the likelihood of end users encountering failures.
What if an end user reports an error but no rejection was seen during testing? The generic error message on mobile (see Section 3b) can appear for photo quality issues as well as safety rejections. If end users are reporting failures, ask them to confirm their photo is well-lit and shows a clear, unobstructed view of their face. If the photo appears valid and failures persist, escalate to your CUE support contact.
5. Prompt Writing Recommendations
Avoid ambiguous costume or appearance descriptions
Be specific and conservative when describing costumes and outfits. Instead of implying revealing clothing, describe it explicitly as performance wear, theatrical costume, or fully clothed stage attire.
Use explicit modesty and context language
Adding phrases like the following can reduce rejection rates:
“fully clothed in a theatrical costume”
“PG-rated, family-friendly event attire”
“modest, stage-appropriate outfit”
“vintage Las Vegas showstopper look, fully costumed with feathers and sequins”
Be direct about the setting
Clearly frame the scene as a performance, event, or theatrical context to reduce ambiguous interpretation. For themes that could imply violence (e.g., zombie, warrior), emphasize the theatrical or Halloween-costume nature explicitly.
Test with diverse subjects
If possible, always generate a preview with at least one male and one female test image before publishing. A prompt that passes with one may fail with the other depending on how OpenAI evaluates the combination.
6. If a Prompt Cannot Be Fixed
If prompt adjustments do not resolve the rejection:
Note the Request ID from the rejection message (format: req_[alphanumeric]).
Submit a request to OpenAI at help.openai.com, including the Request ID and a description of the intended use case.
In the meantime, work with your Client Success representative to find an alternative prompt that achieves a similar visual result within OpenAI’s guidelines.
