I've been integrating Midjourney into a content generation pipeline for a prototype UI mockup tool. The goal was to rapidly produce placeholder images for various business domains (e-commerce, fintech, health dashboards). The experience has been technically fascinating but operationally frustrating due to the black-box moderation system.
The filters seem to trigger on an extremely broad and unpredictable set of prompts. For example:
* Prompt: `"a graph showing stock market growth, red and green arrows, professional dashboard"` – **Flagged.**
* Prompt: `"a person holding a smartphone with a banking app interface visible, blurred background"` – **Flagged.**
* Prompt: `"an abstract background of connected dots and lines, representing a secure network, blue tone"` – **Allowed.**
Attempting to generate images for a hypothetical pharmaceutical company (e.g., `"a happy senior patient in a clean waiting room"`) was nearly impossible. The system appears to block any prompt containing certain medical-adjacent keywords, regardless of context. This makes it unusable for professional sectors like healthcare, finance, or even education where any human representation might be construed as "risk."
From an API design perspective, this creates several critical issues:
* **No Error Specificity:** You get a generic "flagged" message with no detail on which term or concept triggered the filter. This makes iterative prompt engineering a guessing game.
* **Inconsistent Boundaries:** The same prompt may pass or fail at different times, making any automated or CI/CD-integrated workflow unreliable.
* **No Allow-listing:** For enterprise or professional teams, there's no way to pre-approve a domain or set of safe keywords, forcing all work into a public channel.
The lack of transparency and control is a significant barrier to adoption for professional, compliance-conscious use cases. It feels like using a powerful API where 30% of requests randomly return a `403 FORBIDDEN` with no `X-Reason` header.
benchmark or bust
benchmark or bust