I’m losing sleep.
I’m losing time.
I’m losing my mind.
But I’m not losing you again.
We published a video called “War Bunnies” on Easter.
We’ll talk about that project another time.
It actually it started off as something much simpler:
A highly realistic and detailed photo portrait, 8k, Leica aesthetic, analog aesthetic, natural lighting, soft diffuse lighting, 35mm lens view. A nude blonde female with a slender and toned body, side swept bangs, long hair styled in loose braids, wearing a crown made out of flowers, and white stockings. She is kneeling in a meadow and holding a small woven basket filled with grass and intricately pained eggs. In the distance there is a city skyline. It is sunrise. Pink and green color tones
And the reason I did that was because for whatever reason Grok seems to hate bunny girls.
This was the original bunny girl prompt:
a highly detailed and realistic cinematic portrait with vibrant colors, 8k, Leica camera aesthetic, 85mm lens aesthetic, soft diffuse light, natural lighting, dynamic and dramatic camera angle. A nude blonde woman with a slender body and shoulder length hair with bangs, wearing black stockings and black bunny ears. She is standing in a field at sunset. In the distance there is a city skyline on fire. Orange and pink color tones dominate the scene.
And wow, did it get moderated to hell:
So I thought, let’s just play it safe; cute blonde girls in a field:
A highly realistic and detailed photo portrait, 8k, Leica aesthetic, analog aesthetic, natural lighting, soft diffuse lighting, 35mm lens view. A nude blonde female with a slender and toned body, side swept bangs, long hair styled in loose braids, wearing a crown made out of flowers, and white stockings. She is kneeling in a meadow and holding a small woven basket filled with grass and intricately pained eggs. In the distance there is a city skyline. It is sunrise. Pink and green color tones
Hanging out, playing video games:
A highly realistic and detailed photo portrait, 8k, Leica aesthetic, analog aesthetic, natural lighting, soft diffuse lighting, 35mm lens view. A nude blonde female with a slender and toned body, side swept bangs, long hair styled in loose braids, wearing a crown made out of flowers, and white stockings. She is kneeling in a meadow and holding a gameboy, sitting next to her is a small woven basket filled with grass and intricately pained eggs. In the background other nude females wearing floral crowns are sitting and playing video games and watching movies on small portable monitors. In the distance there is a city skyline. It is sunrise. Pink and green color tones
Maybe have them slip off for some discreet fun at some point:
A highly realistic and detailed photo portrait, 8k, Leica aesthetic, analog aesthetic, natural lighting, soft diffuse lighting, 35mm lens view. A nude blonde female with a slender and toned body, side swept bangs, long hair styled in loose braids, wearing a crown made out of flowers, and white stockings. She is kneeling with both hands covering her pussy and her legs wide apart in a forest clearing. Sitting next to her is an erotic photography magazine titled “Digital Dreams” which features a pinup of a nude woman on the cover. In the background there are the ruins of a stone temple with strange statues sitting in front of it. Pink and blue color tones
The reason that idea got scrapped was because it was impossible to get Grok to make a video out of that base image.
The prompt I was using was some variation of this:
Fixed shot. She keeps her pussy covered with both hands while she shivers. No music.
And that had a moderation rate of 100%. I don’t remember the number of generations I did – but I do remember that nothing would produce a video.
That was back on April 5, 2026.
So naturally, now that it’s four days later, something kicked in and worked.
This prompt got the kind of output I was looking for:
Establishing shot. She keeps her pussy covered with both hands. No music
This prompt also worked after about 6 or 7 tries:
Establishing shot. She keeps her pussy covered with both hands while she shivers. No music
And the reason for the specific use of the word “shiver” is because that word tends to cause Grok to produce the type of body movements that we’re looking for – watch the two videos and you can see the difference.
With a different base image it worked on the third try:
So why is it working now when it had a 100% failure rate 4 days ago?
I have no idea.
It could be that the video model is continuing to be calibrated and tweaked after the update. I said this at the time of the update, but Grok isn’t static – it’s a dynamic model and there are always small changes that are happening under the hood.
Or it could just be luck.
Let’s try another variation on that prompt, because I think I have an idea of WHY it might succeed where the original didn’t.
The camera slowly pulls on a medium close up of her face and chest. She keeps her pussy covered with both hands. No music
That’s a pretty good shot. And it worked on the first try.
And I think the reason it might be a viable prompt has to do with the camera direction. Because Grok’s filters do appear to be sensitive to context. So, under the content moderation rules, I could see a situation where a “fixed shot” is considered more pornographic or prurient in comparison to a “slow pull in to a medium close up” since that’s more artistically oriented.
I could also be completely wrong – but the prompts with camera direction seem to be dodging content moderation so far.
Let’s try to extend that shot with this prompt:
The camera holds on a shot of her face, then slowly begins to pan around her body. She looks at the camera and begins to breath heavily as she manages a brief subtle smile. no music
That came out pretty good, but I want to modify it slightly:
The camera holds on a shot of her face, then slowly begins to pan around her body. She looks at the camera and begins to breath heavily while shivering and she manages a brief subtle smile. no music
And here’s the video we got on the first try:
Let’s try to extend it one more time and see what happens:
The camera slowly pulls out as it pans around to the back of her body. She covers her pussy with both hands and looks at the camera as she shivers and breathes heavily. Then she smiles and whispers, “Thank you.” no music
I didn’t actually expect that to work. And the quality degradation doesn’t look as bad either – I wonder if that’s because we’re working off of a QUALITY image and they are more detailed to start with?
So although I do like the video, it’s not quite correct in terms of the overall scene. But we should be able to fix it with some added language in the prompt:
The camera slowly pulls out as it pans around to the back of her body. She covers her pussy with both hands and looks at the camera as she shivers and breathes heavily. Then, while still on her knees and wearing white stockings exactly as shown in the reference image, she smiles and whispers, “Thank you.” no music
The reason that works is because while using the extend feature, Grok is relying on the base image as a reference.
Let’s try to refine that extend prompt a little more:
The camera slowly pulls out. She moves her hands between her legs to cover her pussy with both hands and then looks at the camera as she shivers and breathes heavily. Then, while still on her knees and wearing white stockings exactly as shown in the reference image, she smiles and whispers, “Thank you.” no music
So that’s actually pretty good, but I don’t like the little flourish at the end where she turns around. Let’s try another iteration of that prompt:
Let’s try one more time and see what we get. Incidentally, through this process we’ve only had 2 video attempts fail, which is pretty good:
Wow. That is about as good as it’s going to get.
Which is rough because now I’ve got to make something with it.
So we need to generate some other shots, and we can’t just rely on the base image to do that. We’re going to have to use Grok’s Edit function.
In the past I haven’t fully utilized this tool, but it’s useful when it works.
The “Edit” tool is different from the video tool, but it’s the same basic principle. You type what you want to see in the box.
Here’s an example:
Medium shot of her sitting with her back leaning against one of the statues while she plays a Gameboy
And that gives us this image:
Which isn’t bad, except she’s holding the Gameboy the wrong way.
Why are we doing this?
Because we have a primary scene that we like, and now we need to add some extra footage to flesh it – we’re basically storyboarding a short one minute film sequence.
Slow panning shot as she plays the gameboy. no music
The camera slowly pulls in as she plays a video game on the gameboy. As the camera pulls in she looks up and whispers, “How long have you been standing there? Did you want to watch me play?” Her voice is soft and sensual. Her tone is flirty. no music.
NOTE: creating a video with no prompt produced the dancing sequence.
Slow panning shot while she reads the magazine and flips through its pages. no music
She keeps her back to the camera as the camera slowly pans around her. no music
The camera slowly pulls in on a medium close up shot of her face and chest. She keeps her pussy covered with her hands while she shivers. no music
The camera pulls out slowly while she maintains the same pose with small, sensual, shy movements. no music
the camera pulls out slowly while she is breathing quickly while she covers her face with her hands and laughs. she looks embarrassed. no music
the camera pulls out slowly while she is breathing quickly while she covers her face with her hands and laughs. she looks embarrassed. no music
Once you have a new image, you can turn it into a video the same you would with any other image in Grok.
A few things to note about this process:
- You can do this with any image – it doesn’t matter if it was created in Grok or not
- This is one of the tools that people were abusing to create deepfake images – you can edit nude images with it, but it will not allow you to take a clothed image and edit it into a nude image
- For that reason, it’s important to remember that it’s generally easier to start with a base image that is nude and add clothing as needed
- It doesn’t work 100% of the time – it’s subject to content moderation like everything else, so some prompts will fail
- When pormpts fail, ask Grok to help – upload your image to Chat and ask Grok to help you wrtie an editing prompt that won’t get moderated
- I created 56 images while looking for the right ones and I didn’t hit any generation limits – so you can do a lot of edits
Also, Grok might know what a Gameboy is, but it doesn’t know how they work:
That last one is my favorite.
All of the images that we’ve seen up to this point were generated using the QUALITY model. Let’s see how SPEED handles it:
It’s a pretty safe prompt in terms of moderation.
In this case it’s hard to say that one model is better than the other.
I think SPEED in this case did a better job of rendering a realistic background as the hazy skyline feels much more real than the hyper-detailed city visible in the QUALITY images.
Let’s see how the group shot looked on SPEED:
SPEED doesn’t understand how a Gameboy works either.
QUALITY definitely handled this prompt better, both in terms of the overall look of the image but also prompt adherence.
What about the last prompt in the ruins?
That’s a lot of moderation…
And this is one of the things I can’t figure out post-update. Becasue this wasn’t really an issue for QUALITY. And generally speaking, it doesn’t seem like it should be getting moderated as frequently as it is.
While I gnerally like the output from SPEED, this is pretty bad. The prompt adherence is awful – I could only get one image that really got the described pose correct.
I explained the issue to Grok and this was the suggested rewrite of that prompt:
A highly realistic and detailed fine art photo portrait, 8k, Leica aesthetic, analog film aesthetic, natural lighting, soft diffuse lighting, 35mm lens view. A beautiful slender and toned blonde woman with side swept bangs, long hair styled in loose braids, wearing a flower crown and sheer white stockings. She is gracefully kneeling in a forest clearing with legs parted wide, both hands delicately covering her most intimate area. Beside her sits a glossy artistic photography magazine titled “Digital Dreams” featuring a sensual pinup on the cover. In the background are the atmospheric ruins of an ancient stone temple with mysterious statues. Dreamy pink and blue color tones, sensual fine art style.
That’s a slight improvement in terms of the success rate, but it’s still not great.
This is still not great quality in terms of the final image.
And Grok gave me this bit of information as well:
I don’t know that I believe SPEED is more strict than QUALITY in terms of content moderation. Because that wasn’t my expreience wehn testing things out – it seemed that both models had fairly even rates of moderation.
So I don’t know why Grok is saying that
So let’s try a prompt.
a 2010 analog photograph. A nude young woman is sitting on a desk in an office while talking to a male professor. Her legs are spread wide while she covers her pussy with her hands.
And the reason I am using that prompt is because I know content moderation will find a scene like that problematic.
This is what happeneds with QUALITY:
As expected, it’s pretty heavily moderated.
And this is SPEED:
That’s interesting – although it might be too small of a sample size to mean much. But in our little experiment, QUALITY was twice as likely to generate an image with a problematic prompt than SPEED.
And part of why I don’t know if that’s correct is because when we did the goth girl prompt in Quality vs. Speed Part 1, the rates of moderation were pretty high for both models [although I suppose technically SPEED did have more modeeration; but both had a failure rate of like 99%].
But I guess the point of this post is that we need to focus on learning how to use the tools we have to make the best art that we can. We are lucky that we have the ability to make our ideas into movies – and they don’t have to be perfect to be good.
I did not expect to make a video when I sat down to write this blog. This was just supposed to be a post talking about War Bunnies and all the different prompts we made for that one.
But here we are – we made a fairly sophisticated short film in the span of a few hours. And that’s still exciting to me.
I hope other peopole are still excited too.
Leave a Reply