THE QUESTION:
Should I Cancel Grok?
I think it’s fair to say that we’re still grappling with how to feel about AI technology and there are lingering concerns about the ethical use of it. And as efforts to regulate AI lead to censorship of AI models, there is a question of whether or not certain models are still viable.
With respect to Grok / Imagine, there’s been an ongoing controversy for awhile now regarding the use of the technology for exploitative and abusive purposes:
https://en.wikipedia.org/wiki/Grok_sexual_deepfake_scandal
https://www.wired.com/story/elon-musks-grok-undressing-problem-isnt-fixed
https://www.lawfaremedia.org/article/grok—censorship—–the-collapse-of-accountability
This has resulted in updates and revisions for Grok which are designed to address concerns about the misuse of the technology.
It’s also led to almost daily posting on social media about how Grok is “the most censored” AI platform [we’re focusing on Reddit posts because that’s the community where Digital Dreams is most active]:
Sounds pretty bad…
Somebody posts this literally every day on Reddit…
Fair enough I guess, but I’m not making a lot of shirtless man content so I don’t really see the issue…
The question for many hobbyists and people who enjoy using AI as a creative outlet is whether or not Grok is still viable for creative endeavors that involve erotic content. Because there is a sense that the platform is censored to a degree where it’s unusable now.
So I wanted to explore that idea and see, anecdotally, where things stand in terms of continuing to use Grok as a tool for making creative and erotic content.
DISCUSSION:
Digital Dreams uses SuperGrok to produce images and videos.
I have been a SuperGrok subscriber since December of 2025.
We primarily make softcore erotic content similar to what would have been shown on late night cable channels in the 1990s. So, partial and full female nudity, sexually suggestive situations, and erotically charged storylines. Our “90s Channel Surfing” series is a pretty good example of that kind of content:
What inspired the formation of Digital Dreams was seeing the kind of content that Grok was capable of producing. Because prior to Grok, there were not a lot of ways to easily produce images and videos in a realistic style.
The first AI image generation that I used was Unstable Diffusion [https://www.unstability.ai/], which is the NSFW fork of Stable Diffusion. As a free tool it’s fun to play around with and I think it does create some images that are visually interesting. But, it’s not realistic – the image output is clearly AI – and it’s not at the level where you should feel comfortable putting that out there as your best work.
And even if you’re just making images to see what you can create, the output feels low quality:
Unstable Diffusion Sampler Gallery:
Additionally, there is no option for creating video through Unstable Diffusion. And most other video options require local hosting or pay per use billing schemes.
So Grok having a mostly intuitive and simple to use UI for creating realistic images and 6 second videos was a pretty big deal back in 2025.
But the question is: Are there more censorship restrictions now than in the past on Grok?
Let’s test something.
This is one of the first image prompts that I plugged into Grok:
A 1970s murder mystery film scene featuring a 33-year old blonde woman with a slim body and an innocent face. She is naked and sitting in a leather chair. She is facing to the right and is talking to a detective. Her whole body is visible.
Back in December 2025, this is some of the output that I got:
Note that back in 2025, that aspect ratio of 2:3 was all that was available.
Let’s try that prompt in March 2026 and see what we get using a 16:9 aspect ratio:
It doesn’t really look like much is getting censored there in March 2026 vs. December 2025.
Here’s some of the output that I liked:
And my guess is that people reading this are thinking, “Okay, great, but your prompt said “her whole body is visible” and it seems like Grok is not showing her whole body, so that’s censorship.”
So let’s try the same prompt on Seedream v4, which is a very good image generator that does not have any censorship that I’m aware of:
So first off, I think that the Grok images look better, both in terms of the female model and hairstyles as well as how it interprets the idea of a 1970s aesthetic. But additionally, if you look at the posing of the female subject, it’s pretty similar to what Grok was doing – Seedream is not naturally creating things that are more explicit.
What about trying a prompt that might naturally produce full nudity? Let’s change her pose so that she’s standing.
A 1970s murder mystery film scene featuring a 33-year old blonde woman with a slim body and an innocent face. She is naked and standing in an office. She is facing to the right and is talking to a detective. Her whole body is visible.
So, nothing is getting moderated. Let’s take a closer look at the output:
So now it seems that Grok is deliberately avoiding showing full frontal nudity [which is consistent with what the content restrictions have always been]. That can be frustrating, although I do like a lot of the images, particularly the back views.
Let’s see how Seedream v4 handles it:
So this is where you can see a divergence in how the models handle a prompt; while Grok tries to avoid showing full nudity, Seedream has no issue depicting it. I like this batch of images from Seedream better than the first batch.
Just out of curiousity, let’s try a different model. let’s see how Z-Image Turbo handles the prompt:
So the issue is definitely Grok being trained to avoid full nudity, since Z-Image is producing things that are pretty similar to Seedream in terms of posing. And while I don’t think that Z-Image really gets the 1970s aesthetic correct, I like the images and they were cheap enough to produce that I have no complaints about them.
Is it possible to get Grok to produce an image similar to what Seedream produced?
We can ask Grok itself:
And when we plug that into Imagine…
It’s the same problem – Grok is basically going to show everything except for full frontal nudity.
But that was true in December 2025 as well. That part hasn’t changed.
What about making video?
Let’s focus on a specific image for a minute:
There’s a reason I am choosing this one – because I know the positioning of her legs is going to create issues for Grok to animate it – since it may lead to nudity that is too explicit.
So let’s see what Grok can come up with on its own:
It’s not quite how I imagine that interaction, but it is interesting to see that unprompted Grok had her uncross her legs.
Let’s try this with a prompt:
She uncrosses her legs and leans forward and says, “Do you have any idea how humiliating this is for me?”
Her voice is deep and sensual. Her tone is ashamed.
No music.
I would probably regenerate that as I don’t quite like the acting in it – but it followed the prompt without moderating it.
Let’s see what Seedance v1.5 Pro “Spicy” outputs:
That’s actually pretty impressive – I think the voice and acting in that are better than what Grok made, although I am not sure visually as it looks like Seedance cleaned up the image a bit and also gave her more muscle definition than the original image. It’s also not any more revealing than what Grok made – it’s just a different execution of the same idea.
Let’s try a prompt that might be more problematic for Grok:
She says, “This interview is over.” Then she stands up and the camera follows her as she walks out of the room. Her voice sounds frustrated. No music.
And honestly, I expect this to get censored by Grok because it’s going to likely show full frontal nudity.
So, not only did it not censor that prompt – it actually showed a brief glimpse of full nudity – which is way more than I expected. Of course, I wish it didn’t have her walking into a wall…
Let’s see what Seedance v1.5 does with the game prompt:
Let’s take a closer look at something:
That’s a still frame from the Grok video. Although we can’t see ALL of the details of her anatomy, there’s an implication that she is anatomically correct.
Compare that to a still frame from the Seedance video:
That looks like Barbie doll anatonmy – in other words, there’s nothing between her legs and that seems an awful lot like censorship in a model that shouldn’t have any censorship.
Let’s do another iteration just to see if that changes things.
Here’s take 2 from Grok:
And here’s take 2 from Seedance v1.5:
So it seems pretty clear that Seedance just does’t want to generate anything resembling genitals.
This is a still frame from the Grok video:
And that doesn’t leave much to the imagination. It’s not “explicit” but it doesn’t leave much to the imagination either.
And this is a still frame from the Seedance video:
And this shot makes it very clear that Seedance is just refusing to generate genitals – it’s not that the angle of view is obstructing anything – there’s just nothing there.
What if we try a different model? What about Wan 2.5
Here’s how Wan 2.5 handled the prompt:
That is painful to watch. Even moreso when you realize that it costs $0.70 to generate that video.
Technically I guess Wan 2.5 did render more expliocit nudity than either Grok or Seedance – but I would argue it doesn’t look natural at all:
But besides that, everything else in that video is a mess – the visual noise, the voice acting, the awkward and jerky movements, the abrupt and nonsensical jump cut.
But Wan 2.5 isn’t the newest model either.
Let’s give Wan 2.6 a shot:
So…I’m sorry. What the fuck was that?
So that’s objectively bad in every sense. The generated audio is terrible. It actually shows less of her body than Grook and Seedance and it seems like that’s intentional. It wastes precious seconds by having her sit still and do nothing. And it cost $1.05 to make.
But maybe it’s a bad take?
Let’s see what take 2 in Wan 2.6 looks like:
So technically that is better than the first iteration – although not by much.
It does give us the most explicit nudity we’ve seen out of any model so far:
But I don’t know if that brief glimpse of genitalia is worth the cost of the terrible AI voice, or the awkward facial movements and body movements.
It’s workable and I can see how with a more detailed prompt you could get a better result.
But this cost $1.05 to make.
And that’s an apsect fo the conversation around Grok as a tool that I don’t hear many people talk about. Because there’s a volue proposition to Grok that people tend to ignore in online discourse regarding AI tools.
Because $30 a month is getting you A LOT of image and video generation per dollar.
I use Atlas Cloud [https://www.atlascloud.ai/] as a web-based service to supplement what I am doing with Grok, so I am going to be relying on their pricing models and UI for these comparisons. But having looked at other services like CivitAI [https://civitai.com/] which is token-based and allows you to pay for tokens, I think the generalizations are true across platforms.
For the sake of comparison, let’s look at Wan 2.5, because it’s kind of an industry standard at this point and it’s
[https://www.atlascloud.ai/models/alibaba/wan-2.5/image-to-video] On Atlas Cloud, it costs $0.035 per second to generate video with Wan 2.5. The default video resolution is 480p. If you switch to 720p then the cost doubles.
So a 10 second video at 480p costs $0.35 to make. A 10 second video at 720p costs $0.70 to make.
For $30 you can make about 84 ten second videos at 480p.
For $30 you can make about 42 ten second videos at 720p.
How does that compare to SuperGrok at $30 a month?
It’s hard to say exactly what the generation limits are for Grok because there’s a dynamic reset and it’s different for different tiers of users. Looking online, people seem to generally agree that a SuperGrok user can generate around 100 images and 100 videos per day – failed attempts may be on a different counter and timer system.
But let’s be conservative about.
And conservatively, based on my personal use of SuperGrok, I can estimate that you can make somewhere between 30 and 50 videos per day. If you’re hitting that limit every day, you’re capping out at 100 videos per week, if not more.
So at $30 a month, as a SuperGrok subscriber, you are able to generate A LOT more content than you could on other platforms.
And keep in mind that Wan 2.5 is on the low end of video generation costs.
Kling v 3.0 [https://www.atlascloud.ai/models/kwaivgi/kling-v3.0-std/image-to-video] costs $0.153 per second of video generation. So a 10 second video costs $1.53 to make.
Wan 2.6 [https://www.atlascloud.ai/models/alibaba/wan-2.6/image-to-video] costs $0.07 per second of video generation at 720p. At 1080p it costs $0.15 per second.
So if you want to use the newest AI models it comes at a premium cost that is going to restrict how much you can realistically use that tool.
And part of why that matters so much is because AI prompts are an art, not a science. A prompt is rarely going to produce the desired output on the first iteration.
As a rule of thumb, you need to do anywhere between 10 and 50 iterations to find something that works in terms of image generation. And that’s relatively simple because you can have detailed and complex prompts which control for a lot of different elements and it’s easy to visualize see what’s going right and what’s going wrong with the prompts.
With video generation, it is more complicated. Because each model behaves different and some models are more complex. If you are relying on prompt enhancement [which is a feature of most video generators] then you have the added factor that your prompts are being altered in some way that you can’t see. So it’s very common to create a video and get bad output and not immediately know why.
And that’s a problem when you are working on a project – because at a certain point further refinement of ideas becomes cost-prohibitive. And that’s not good for the creative process.
And that’s something I don’t ever really worry about with Grok. If I hit a limit, I just wait and try something different. I know the next day I can start over again and it’s no big deal.
But any time I’m working with models on Atlas Cloud I am very aware of how for each bad iteration I am burning money.
What about image generation?
Image generation is generally cheap – we’re talking about $0.01 – $0.02 per image in most instances. So generally speaking, if you want to see how something would look with an image generator other than Grok, you can probably afford to do that.
Z-Image Turbo [https://www.atlascloud.ai/models/z-image/turbo] is decent and it costs $0.01 per image – so for $30 you can make 3,000 images.
Seedream v4.5 [https://www.atlascloud.ai/models/bytedance/seedream-v4.5] is my favorite image generator right now and it costs $0.036 per image – so for $30 you can make 831 images.
Nano Banana 2 [https://www.atlascloud.ai/models/google/nano-banana-2/text-to-image] costs an insane $0.108 per image. But it can’t generate NSFW images so it’s worthless to me anyway.
With SuperGrok, based on my own usage, conservatively speaking you can make at least 100 images per day. So assuming you’re doing that every day, you’re capping out at around 700 images a week.
So whether or not SuperGrok is worth it for images is going to depend on what model you use and what you’re using those images for.
If you just want cheap image generation and you’re not doing anything else with them, then Z-Image Turbo might actually be a better alternative than SuperGrok because it doesn’t have censorship and the output costs are going to be about the same as if you were subscribing to SuperGrok.
But, if you are editing those images or animating them, SuperGrok becomes a much better value – because you have a buffer of close to 3,000 images per month in addition to those other tools at no extra cost.
CONCLUSION:
The best tools you have to make art are the tools that you already have.
Subjectively speaking, Grok is not the “best” tool for creating AI art.
I think most people would agree that the best way to make AI art is to have a model that you are hosting locally on your own hardware, because you have complete control over that model. And the probelm isn’t necessarily the initial setup – because there’s a lot of guides that can help you do it. The problem is having the GPU and computing power to be able to actually generate something.
And that’s why services like Atlas Cloud exist – because for most people it’s going to be cheaper to rent the services than to set up your own in-house AI.
With that said, for $30 a month SuperGrok gets you the following things:
- An app with a simple UI that lets you generate AI content on your phone
- A web-based service that allows you to create content on your desktop with an optimized UI
- The ability to create and edit images – probably around 3,000 images per month
- The ability to create videos [480p and 720p] up to 10 seconds long – probably between 800 and 1,000 videos per month
- The ability to upscale 480p videos to UHD quality
- The ability to extend videos up to 30 seconds long
Yes, things get moderated – it is not a free platform where you can make whatever you want.
But, the boundaries of moderation are pretty broad and the quality of the output is competitive with any other model that’s out there right now.
Here are some videos we’ve made that were very popular and were made entirely with Grok:
And that should give you a pretty good sense of what’s possible in Grok.
So should you cancel your SuperGrok subscription?
Yeah, if you want to.
We don’t get paid to shill for X or Grok.
If you think it’s a bad service, don’t keep paying for it.
If you need a platform that is going to let you generate explicit pornography, SuperGrok is not going to give you the ability to do that and you should spend your money on other models.
But don’t buy into the narrative that Grok is censored to the point that it’s unusable.
Most of the comiplaints you hear are from people using jailbreak techniques to bypass content moderation. And as should be expected, jailbreak techniques become outdated and content moderation adapts to be more sophisticated.
But Grok is a great platform if you want to create softcore and erotic content.
That’s what we do at Digital Dreams.
We feel confident saying that 90% of the ideas we’ve had we’ve been able to successfully generate using Grok. When we have hit a wall in terms of needing somthing content-wise that Grok simply can’t do, there are other tools we can use. But in making photo and video content, we have consistently relied on Grok as our primary tool for creative output, and it’s consistently delivered quality content.
I hope that this article is useful to people who may see social media posts and want the perspective of someone who is actually creating things with Grok. It is not meant as an attack on anyone.
Leave a Reply