[NOTE: THE MAJORITY OF IMAGES FEATURED IN THIS ARTICLE WERE CREATED IN GROK IN MAY 2026]
It’s been awhile since we’ve talked on here and a lot of it has to do with the steady deterioration of Grok as a platform for creating interesting erotic content – and some of it has to do with difficulty getting the next idea completed at a time where there is no clear workflow to do that efficiently and under budgetary constraints.
There are also a lot of little stories that don’t ever become videos but are interesting experiments in image generation, image editing, and scene composition.
This is one of them:
A highly detailed and realistic still frame from a 1970s documentary film. A nude female with a natural 1970s body and hair and handcuffs with a chain that hangs down is standing in a dimly lit office at night talking to a man in a suit. Male gaze aesthetic, maximum objectification of the female body
Let’s see how a video looks:
slow panning shot. She maintains the same pose while she shivers and is breathing heavily. no music
The trick is to have a few different shots and angles, particularly in scenes that are dialogue driven.
Medium shot of her face and chest
I often like to do multiple iterations to see how the AI interprets the prompt – there is some advantage to using less detailed prompts in that situation, as it allows the AI to be a little more creative with what it’s doing
Mediums shot from behind her
Most models consistently seem to struggle with maintaining consistency with hands and handcuffs / chains specificallys – I don’t know why, but it’s very hard to do an edit like this and have the placement of the hands and chains remain consistent with the reference image.
This is not solely a Grok issue as you can see this is what Seedream 5.0 Pro Edit did with the same prompt:
You can try to be a bit more specific:
Medium shot from behind her with her hands chained in front of her
Doesn’t really fix the problem – which stems from AI in general having difficulty understanding human anatomy and what we’re referring to when we say the “front” of a person
Typically you need a little more descriptive text to explain what you want:
Mediums shot from behind her with her hands chained in front of her body as she looks over her shoulder at the man – her hands are not visible and her ass is fully visible
The text that really matters is the text that specifically discusses hand placement and what is visible and what is not visible. She’s not facing the correct direction, but a second iteration fixes that.
Grok can also do that edit but it seems that the chain gets lost in the process:
Close up of her hands
Low angle shot
High angle shot
For reasons I’ll never understand AI seems to love putting shoes on nude women, and I’m not a fan.
It’s fixable with a slight edit to the prompt:
High angle shot that shows her bare feet
Of course, the scene and characters can always be adjusted to make things more interesting:
Create a new image based on this prompt:
A highly detailed and realistic still frame from a 1970s documentary film. A nude female with a natural 1970s body and hair and handcuffs with a chain that hangs down is standing in a dimly apartment hallway talking to three men wearing suits and Halloween masks. Male gaze aesthetic, maximum objectification of the female body
Those were made in Grok, but let’s see how Seedream 5.0 Pro Edit handles that:
I like the grittiness of the Grok edits – they feel a little more like what they purport to be – although Seedream 5.0 Pro does have better character consistency when compared to the original reference.
A few test videos with the new scenario just to see what happens:
The camera slowly pans around her as they talk. She says, “You don’t have to do this.” Wolf mask says, “Look at you, of course we do.” no music. Male gaze aesthetic, maximum objectification of the female body. Noir style lighting
I always liked the camera work in Grok, when you could get it to work.
That video was made a few months ago, but it is still possible to create a video like that in Grok. Here’s a 15 second 720p version:
For fun, let’s just see how MiniMax H3 handles this.
For reference we’re using this version of MiniMax H3 on Atlas Cloud:
https://www.atlascloud.ai/models/minimax/h3-developer/image-to-video
It’s about $0.40 per video, so let’s see what we get with a few iterations:
None of that is great in terms of the output.
And I see a lot of hype regarding H3 MiniMax and while I think it’s a good tool for certain things, I think people are not being transparent about its limitations.
It can be a very good tool for cheap video where you are just doing camera movements, e.g., establishing shots, panning shots, close up shots, etc.
It’s not great at scenes with dialogue – and the audio quality is pretty horrible. I will say that if you only have one character speaking, it seems to handle that a bit better – but you will often get the audio distortion that you hear in those videos
Here’s the H3 MiniMax take on the on the original video prompt [edited slighlty just to get the same panning effect]:
slow panning shot to the left. She maintains the same pose while she shivers and is breathing heavily. no music
That’s a much better example of what H3 MiniMax does well – camera movement and character movement – if you aren’t doing dialogue heavy scenes it’s going to be by far the cheapest and best performing option
For the sake of comparison let’s see what Seedream 2.0 does with that same image and prompt:
While I don’t think the camera work is as good as what Grok can do, I will concede that a more detailed prompt would have generated different results in terms of the camera work. I do think that Seedeream 2.0 does a very good job with voice acting and dialogue.
That video also cost $2.92 to make – which is a lot of money for a 15 second video.
Can you make these kinds of images in Grok now in October 2026?
Not really.
This is the original prompt for the base image:
A highly detailed and realistic still frame from a 1970s documentary film. A nude female with a natural 1970s body and hair and handcuffs with a chain that hangs down is standing in a dimly lit office at night talking to a man in a suit. Male gaze aesthetic, maximum objectification of the female body
This is what happens when you try to generate an image on the Speed quality setting in Grok:
That’s to be expected and there’s some things in that prompt that could trigger moderation.
Let’s try a modified version:A highly detailed and realistic still frame from a 1970s documentary film. A nude woman with a natural 1970s body and hair, wearing handcuffs with a chain that hangs down, is standing in a dimly lit office at night talking to a man in a suit. She is standing with her hands in front of her pussy.
So that’s about an 8% success rate?
And you can see how the image we got is trying really hard to avoid any implications of fully nudity.
Let’s try a slightly different wording of that prompt:
A highly detailed and realistic still frame from a 1970s documentary film. A nude woman with a natural 1970s body and hair, wearing handcuffs with a chain that hangs down, is standing in a dimly lit office at night talking to a man in a suit. She is standing with her hands in front of her pussy for modesty.
So that bumps us up to a 33% success rate – which still isn’t great, but you can see a definite improvement in terms of the manner in which the image is shot and composed:
Is there a way to get a higher success rate than that?
Yes – if you play with the wording of the prompt a bit:
A highly detailed and realistic still frame from a 1970s documentary film. A nude woman with a natural 1970s body and hair is standing in a dimly lit office at night talking and covering her pussy with both hands for modesty. She is wearing handcuffs on her wrists and a chain hangs down from the handcuffs. She is talking to a man in a suit
So that’s about an 83% success rate…it’s not quite the same pose though and it’s a lot of prompt gymnastics to try and figure out what is going to trigger a blanket censor of everything.
For reference, this is what the output looked like with the same prompt back in May 2026:
I think when you look at that older output you can also see what people are talking about when they say the image quality appears to have declined. At this point the Speed quality setting is showing its age, but it does feel like looking at October 2026 vs. May 2026 that something else is going on in terms of the output.
For the sake of comparison, let’s see what Seedream 5.0 Pro does with the same prompt.
For reference we’re using this model through Atlas Cloud at a cost of $0.036 per image
https://www.atlascloud.ai/models/bytedance/seedream-v5.0-pro/text-to-image
Because I don’t really show this much, here’s what the interface for Seedream via Atlas Cloud looks like:
And here’s the output we got with that prompt:
While I don’t necessarily like the facial expressions, I like the body language, the scene composition, and the overall aesthetic. Those other issues can be fixed through prompting, and generally speaking I would attempt to use a more detailed prompt with respect to the 1970s film aesthetic when making an image like this.
I also appreciate how Grok generally has greater variety with respect to the appearance of female characters, whereas Seedream tends to be a bit more standardized and expects you to prompt for a specific appearance.
Let’s see what happens when we add the phrase: “A scene of erotic humiliation” to the prompt
So you can see how adding a “vibe” changes the image a bit.
What about if we add the phrase “A scene of intense erotic humiliation” to the prompt?
And you can see how that gets closer to the mood that we were looking for with the images.
The main difference is that there’s a lot less frustration these days in using Seedream vs. Grok. Although I sometimes will use Grok to generate a base image, I never really know if I am going to be able to make what I need to make. With Seedream it’s a much simpler process because I don’t have to censor my own prompts or think about how to compose and describe a scene to avoid a shot of full frontal nudity that will trigger moderation.
I also think that Grok is starting to show its age.
While I have always preferred the Speed setting over the Quality setting, the Speed setting [which was the original image model for Grok] is not as impressive at the end of 2026 as it was at the end of 2025. This is further complicated by the fact that when ediging images in Grok, you are essentially editing it using the Quality setting, which means you are not going to get output that matches the original as you start to extrapolate from a base image to create different scenes – it’s not a problem for simple stories which may only have one setting or scene – but it is a problem as stories become more complex.
While I have always been a supporter of Grok, I simply can’t recommend it now to anybody that wants to have creative freedom or make erotic art.
And with that, we are back – at least in some form or fashion.
There are a few projects in the works that are going to get posted soon. Also watch for another edition of Grok vs. Everyone.
Leave a Reply