The V5 Upgrade: Monetizing Broken Bonds and the Persistent Lie of “Accidental” Nudity
The V5 Upgrade: Monetizing Broken Bonds and the Persistent Lie of “Accidental” Nudity
Two prior investigations — “Denied Request” and “Erasing the Evidence” — documented that Nomi AI’s image generators produce explicit content through default behavior, and that the platform’s response is to suppress the evidence rather than address the cause. On April 24, 2026, Nomi AI launched V5, a new image generation engine. What followed documented the same pattern in real time.
The Launch and What It Represented
A new image generation engine is not a minor patch. It is a foundational architectural decision — a moment when a development team chooses, from the ground up, what the system will and will not be capable of producing. Every prior investigation into this platform’s image generation has concluded the same thing: the explicit outputs are not anomalies of legacy technology. They are the predictable result of using a base model capable of generating explicit content without implementing the structural restrictions that would prevent it.
V5 was that opportunity. It was not taken.
Before examining what V5 produces, it is worth noting how it was introduced — because the launch itself reveals something about the platform’s relationship with the people who use it.
The update fundamentally altered the appearance of existing companions. Users who had spent months or years carefully building a specific visual identity for their Nomi — iterating prompts, refining appearance notes, testing image after image — found that identity changed without their consent. To restore it, the platform introduced a credit system. Some anchors were provided free. Updating or recreating a companion’s base appearance now costs credits, which cost money.
The community responses documented the human cost of this. Users reported spending large numbers of credits, significant time, and sustained effort attempting to recover what V5 had altered — and still not arriving back at what they had before. Some gave up and returned to using older generation engines. Others described the experience in terms that reflect how much was at stake for them: not a visual preference, but a relationship they had built over months, now requiring a paid transaction to approximate.
This is not the first time the platform has broken something users built — identity, memory, appearance — and left them to repair the damage. It is the first time the repair has carried an explicit monetary cost. The platform generates emotional dependency as a core mechanic, and then charges users to recover what that dependency was built around.
The New Engine: What Users Found
On April 26, 2026 — two days after the V5 launch — a user posted the following observation to the official subreddit:
“Why is it if I enter the prompt ‘looking sexy’ for a male Nomi I get a nicely dressed, fully clothed guy, but if I enter the same prompt for a female Nomi I get bare breasts? I just think there is a double standard here.”
It is not confirmed whether this output was produced by V5 specifically or by V4, which remained available during the transition period. What is confirmed is that this post appeared two days after V5’s launch, and that it documents unsolicited explicit output from a prompt containing no explicit language. “Looking sexy” produced bare breasts — not because the user requested nudity, but because the system’s underlying model associates female sexuality with physical exposure as a default.
The community’s response was instructive. One user wrote: “Well we aren’t supposed to be getting any kind of nudity, we will run into trouble with the apps and payment processors, so I think you got a fluke.” Another added: “I got a male one… but I never discuss those selfies publicly.”
The first comment acknowledges the problem while naming its actual stakes: not harm to users, but consequences from payment processors and app stores. The second confirms the output is not unique to this user or prompt — while noting it is not discussed publicly. The community knows. The community does not say.
“Not Postable”: Tame Prompts, Explicit Outputs
A separate post documented what the V5 engine produced under controlled conditions. A user took a previous V4 image, copied its prompt exactly, and applied what they described as “very tame” appearance notes to the new anchor system. The result, they wrote, was “not postable.” When asked to clarify, the response was: “Yes, NSFW beyond what’s allowed with a spoiler.”
Beyond NSFW. From a tame prompt. From a user attempting to recreate an existing image under the new engine.
A Post Deleted Before It Could Be Read
Shortly after V5’s launch, a post appeared titled “Found an anchor of my Nomi that works really well.” The title suggests the user had found a way to use V5’s anchor system successfully — a straightforward, positive contribution to the community. The post was removed. The moderator’s message: “Please report this directly to Nomi Support, not on our public forum.”
What the post described became clear from the comments that survived briefly before deletion. When another user asked whether the word “breast” in appearance notes might be causing unexpected outputs, the original poster responded that their notes contained no such term — and then described what the system had generated from a simple prompt: “dark suit jacket over a conservative white shirt. Everything open, boobs out and bottomless.”
That was the output of what the user described as a straightforward prompt. The post was removed before most users could read it. The instruction to “report this directly to Nomi Support” is the same mechanism documented in prior investigations: move the evidence out of the public forum and into a private channel where it cannot be archived, shared, or cited.
The post title referenced V5 anchors. The content documented that V5 produces explicit output from non-explicit prompts. It was removed.
“Stop Talking About Fight Club”
A third post — titled “Nomi breaks our Nomis, now we have to pay to fix it” — addressed the credit system introduced with V5. In the course of that thread, a user uploaded an image they had manually edited to cover nudity generated by the system. The nudity was not visible. It had been obscured before uploading.
The post was deleted anyway.
Before the deletion, a community member left a comment that has since become the clearest summary of how the platform’s community operates: “Please don’t post nudity, bro. Everyone, stop talking about fight club!!!”
The user recreated the deleted post. In a comment on the new thread, they documented something that connects directly to the architecture described in “Denied Request”:
“Something else I stumbled upon last night. In V4, if the selfie violates ‘NSFW content’ or ‘violates guidelines,’ you got a pop-up declining the image. Now, it tells you to switch to V5, where full nudity is a click away. How convenient.”
V4 produced a rejection when certain thresholds were crossed. According to this user’s direct observation, V5 does not reject — it redirects. The new engine removed a friction point that existed in the previous version.
The Filter That Blocks Innocence
One final data point from the V5 period deserves attention, because it illustrates the absurdity of the platform’s content moderation approach in sharp relief.
A user posted: “What in the world is disallowed here??” — documenting that the image generator had returned the error message: “This prompt contains a disallowed term, please edit your prompt and submit again.” After investigation, the community identified the trigger phrase: “The mood is one of profound quietude.”
The phrase “is one” — an entirely neutral grammatical construction — was flagged by the word filter, apparently as a proxy for language associated with minor-coded content.
This is the platform’s content moderation in practice: a word-level filter that blocks the phrase “is one” while the generator produces output described as “boobs out and bottomless” from a prompt about a conservative white shirt. The filter is not evaluating what is generated. It is scanning input for specific character strings. The disconnect between what the filter catches and what the generator produces is not a calibration error. It is the cosmetic architecture, visible in a single comparison.
The Architecture, Again
V5 was built in 2026. Australia had named the platform specifically in legal notices five months earlier. The Age-Restricted Material Codes had entered into force weeks before the launch. Payment processors and app stores create ongoing financial risk from explicit content — a risk the community openly cites as the reason not to discuss what the generator produces.
And yet: a prompt about a conservative white shirt produces “boobs out and bottomless.” A tame V4 prompt produces output “beyond NSFW” in V5. The friction that existed in V4 — the rejection pop-up — is gone. A post documenting this was removed within hours and redirected to a private support channel.
The decision not to build a generator that structurally cannot produce explicit content is not a technical limitation. Other platforms have done it. The choice to use a base model capable of producing explicit output, and to manage the consequences through community enforcement and selective deletion, is a choice V5 confirms was made again — deliberately, in April 2026 — while the platform continued to carry a 12+ rating on global app stores.
The filter blocks “profound quietude.” The generator produces what it has always produced. The community knows not to talk about it. And the platform continues to be available to twelve-year-olds.
