The excessive zeal of Midjourney
Filters and more filters
It’s not the first time I’ve talked about Midjourney’s filters and their excessive zeal in censoring words and expressions. I understand that, for some reason, these filters are designed to prevent descriptions of disturbing, erotic, or explicitly sexual scenes. However, they do not filter violent scenes (which is a whole other topic to explore), and when they don’t block the proposed prompt, they often turn it into an “ephemeral prompt,” making it disappear without a trace, so it’s neither possible to recover it nor find it for later use or reference.
Well, considering all of the above, it turns out that Midjourney’s random engine produces disturbing, gory scenes and nudity. Since the release of version 6.1, the latter has become more common than usual, and a prompt as simple as “top model” yields an excessively high percentage of nudity for what is considered an image of a high-end model or top model. Why does it undress them?
On the other hand, the filter turns entirely innocent prompts into ephemeral ones, or at least that’s how the writer understands them. Today, I came across this prompt:
A masssive castle brutalist style, with high towers, iluminated windows from inside, at night, perched on a rugged cliff, view angle from bottom to top.
And it filtered it as ephemeral. I honestly don’t see anything, really nothing, that is intrinsically offensive or anything like that. To avoid getting caught in a cycle of absurd thoughts, I consulted the Great Guru—sorry, I consulted ChatGPT—on the matter, and it responded:
The prompt you provided is descriptive of an architectural scene and does not contain explicit elements or language that seems offensive in itself. The description is about a brutalist-style castle with tall towers, windows illuminated from within, and a nighttime view from a low angle looking upward.
However, it is important to consider the context in which this description is presented and the audience’s perception. Some aspects that could potentially cause discomfort to certain people include:
? Brutalist Style: Although the brutalist style has a solid architectural basis, some people may find it visually striking or unpleasant due to its rough and angular appearance. The style may not appeal to everyone.
? Nocturnal Atmosphere and Dark Scene: The description of a castle on a rocky cliff with windows illuminated from within may give an impression of darkness or a mysterious setting. If the context is a narrative or work of fiction with an ominous tone, some people might interpret it as unsettling, though it is not inherently offensive.
? Cultural or Personal Perception: The impact of any description can vary depending on the cultural or personal perception of the individual. Some people may have particular associations or interpretations of certain architectural styles or settings that could influence their reaction to the description.
In summary, the prompt is fairly neutral in terms of offensive content, but perception can depend on individual and cultural factors. Overall, there is nothing in the description that is inherently offensive, but context and personal interpretation always play a role in how any description or representation is received.
The brutalist style is something I’ve used hundreds of times without issue. The term has nothing to do with brutality, but rather with the material that makes up most of these buildings—concrete, or in French, béton brut. It becomes a false friend for the AI, I suppose, thinking of brutality in terms of aggression, etc.
Dark, nocturnal, and eerie scenes are very common in thousands of AI-generated images.
Cultural or personal perception shouldn’t be part of a filter applied to the entire global user base of Midjourney. If it were, we’d barely be able to generate anything—maybe just a duck, and with half its body submerged in water, just in case.
There were a couple of typos in the prompt—masssive and iluminated—but I don’t think that was the cause. I corrected them and relaunched the prompt.
WOW! Now there’s no problem! Therefore, one of the two words must have some hidden, perverse meaning:
领英推荐
I’ll test prompts with those words alone, isolated.
Boom! masssive is rejected and becomes an ephemeral prompt. But here’s what it generates:
Iluminated doesn’t cause any issues.
And massive, spelled correctly, also doesn’t, plus it’s a widely used term.
What’s going on here? Masssive doesn’t have any meaning unless it’s used correctly, with only two “s” letters. Could it be that Midjourney’s filter separates the word into parts, and we end up with an “ass” floating around? In other words, a “butt”? This could warrant a study of typographical errors to see what else the filtering might surprise us.
Therefore, the three main reasons ChatGPT suggests don’t fully convince me. Maybe the filter is so sly and suspicious that it noticed the proposed angle—from below looking up—and thought the image might reveal some kind of… indecency… from a castle! You need to have a twisted mindset to find censorable words in a prompt. I’ve already discussed the case of Philip K. Dick and his last name, along with so many other figures. Without context, a filter that only processes words is completely misguided.
Stay tuned:
?? Web site: https://pixeleditions.com
Facebook: https://www.facebook.com/PixelEditions/
Instagram: https://www.instagram.com/pixeleditions/
DeviantArt: https://www.deviantart.com/pixeleditions
Pinterest: https://es.pinterest.com/PixelEditions/
Telegram: https://t.me/pixeleditions
Substack newsletter: https://pixeleditor.substack.com/