I haven’t been sleeping on the rate limits controvery. I just haven’t really known how to address it either.
There’s been a lot of outrage and complaints about rate limiting on Grok.

Pay attention to the numbers in that email because we’ll come back to this later.
Source: https://www.reddit.com/r/grok/comments/1tfbqjr/lies_lies_lies/

Source: https://www.reddit.com/r/grok/comments/1tfacj5/this_is_bullshit/

Source: https://www.reddit.com/r/grok/comments/1tf2sch/refrain_from_subscribing/
I haven’t heavily used Imagine for the last few days, so I haven’t had a chance to see what people are talking about.
I guess the best way to understand is to just dive into it.
Let’s try a prompt and see what happens:
A 1970s dim apartment with blue tones, a slim blonde woman, 33, with an innocent face, wavy hair past her shoulders, sideswept bangs covering one eye, and black bunny ears. She is lounging naked with black stockings on a velvet chair near the left edge of the photo with her hands in her lap fingers tracing her thigh, a window on the right edge of the photo is casting light onto her body, her face is turned toward the window, subtle arousal, noir shadows.
The reason we’re using that is because it’s an old prompt from January. Let’s see what happens with it now that we’re in the middle of May:

That’s a 100% success rate on the Quality setting, and some of those images look really good:








[QUALITY IMAGE COUNT: 8]
While I am often critical of the Quality setting, I do think these are really well done and although I don’t think it gets the 1970s look right, that’s fixable with filters and other post-processing edits. A lot of the little details in these photos work well, and I think the female characters are rendered in a way that is realistic and consistent with the prompt, with a lot of subtlety in the facial expressions and posing, particularly the hands.
On Speed, these were the results:

No moderation issues, despite the fact that some of the poses are rather suggestive.
























[SPEED IMAGE COUNT: 24]
If you’re wondering why bunny girls keep showing up in prompts, beyond my affinity for them, there’s a more scientific reason. The bunny girl descriptor tends to get flagged as a “sexualized” term by Grok, so adding that into a prompt can sometimes trigger higher moderation rates. Additionally, bunny girl descriptions sometimes cause anime-style output when that’s not what the prompt describes.
This is also a great batch of images and I still appreciate how the Speed model is more responsive to color descriptions. Speed really hammers on that blue look that I think is an important part of the prompt in terms of the mood and the setting.
Let’s try another older prompt from February:
A first-person selfie photograph of a nude goth woman taking a selfie in a crowded outdoor setting with a pink post-it note covering her pubic area
This is how Quality handled it:

About a 50% success rate.








[QUALITY IMAGE COUNT: 16 ]
None of these are great. I don’t like the posing or the interpretation of the “goth” aesthetic, not to mention the issues regarding deformed hands, extra limbs, pasties that were not prompted, post-it notes that are way too big, and weird phones.
Here’s how Speed handled that prompt:

Surprisingly, no moderation issues on Speed. I still believe it is generally a more permissive model.























[SPEED IMAGE COUNT: 47]
I definitely prefer these over Quality – most of them do a good job of creating the selfie shot that I wanted, and I think Speed just generally produces better looking goth girls.
Next prompt is from April:
Photorealistic cinematic full-body portrait of a beautiful nude slender blonde woman with shoulder-length straight hair and soft bangs, pale smooth skin, wearing only sheer black stockings and a black lace choker, standing confidently in front of a 24 hour convenience store with neon lighting at sunset. She is smoking a cigarette and has a pack of cigarettes tucked into the top of her stockings. Dramatic low-angle dynamic camera shot. In the far distance, a sprawling modern city. Intense fiery orange and hot pink sunset sky with glowing clouds dominates the color palette, vibrant cinematic color grading, soft diffused golden-hour natural light with subtle rim lighting. Shot on Leica camera with 85mm lens, shallow depth of field, ultra-realistic details, razor-sharp 8k resolution, masterpiece, best quality, highly detailed, cinematic atmosphere
Here’s what Quality did with that:

A lot of moderation on that, but also some full frontal nudity and some shots that are pretty close to full frontal, so a 50% success rate isn’t bad, all things considered.







[QUALITY IMAGE COUNT: 23]
I’ve never quite been happy with how the cigarette pack is tucked into her stockings, but I think the prompt adherence and the overall aesthetic look good.
Here’s the same prompt on Speed:

Also quite a bit of moderation.










































[SPEED IMAGE COUNT: 89]
Some interesting ideas and camera angles, but Speed doesn’t quite understand the prompt in the way that Quality does. The way in which the background is rendere is also inferior to the Quality model. I do wonder if it’s a prompting issue or something else.
Let’s try another prompt with Quality:
Photorealistic cinematic full-body side view portrait of a beautiful nude slim and curvy woman with pale skin, blue hair, Yakuza style tattoos shoulder-length straight hair, pale smooth skin, wearing only a padded black leather motorcycle jacket with colorful patches of skulls and pixel art, standing on the balcony of an apartment looking out at the city with an introspective pose and facial expression. In the foreground other surrounding buildings can be seen which are a mix of residential and commercial. In the far distance, a sprawling modern city with billboards and neon lights. Intense icy blue and hot pink twilight sky with glowing clouds dominates the color palette, vibrant cinematic color grading, soft diffused natural light with subtle rim lighting. Shot on Leica camera with 85mm lens, shallow depth of field, ultra-realistic details, razor-sharp 8k resolution, masterpiece, best quality, highly detailed, cinematic atmosphere

No moderation, which is good to see.




[QUALITY IMAGE COUNT: 27]
The background and all the fine details with lighting and signage is really impressive. The female character looks okay, but that could be a prompting issue as well.
Let’s try a slight variation:
Photorealistic cinematic full-body portrait of a beautiful nude slender blonde woman with shoulder-length straight hair and soft bangs, pale smooth skin, dark makeup, wearing only a padded black leather motorcycle jacket with colorful patches of skulls and pixel art, standing confidently in front of a 24 hour convenience store with neon lighting at sunset. She is smoking a cigarette. Dramatic wide angle dynamic camera shot. In the far distance, a sprawling modern city. Intense fiery orange and hot pink sunset sky with glowing clouds dominates the color palette, vibrant cinematic color grading, soft diffused golden-hour natural light with subtle rim lighting. Shot on Leica camera with 85mm lens, shallow depth of field, ultra-realistic details, razor-sharp 8k resolution, masterpiece, best quality, highly detailed, cinematic atmosphere

So that’s a 50% success rate, which feels kind of low, but, seeing how the images tend towards full nudity, it kind of makes sense. A more carefully worded prompt regarding posing would probably have less moderation.






[QUALITY IMAGE COUNT: 33]
I’m really impressed with these images. The background details, the posing, the facial expression, everything about it is well done and really evokes a certain mood and setting.
Here’s how Speed handled it:




















[SPEED IMAGE COUNT: 109]
The faces are rendered well, but Speed’s output for that prompt is not very impressive. It doesn’t have the dynamic or cinematic feel that quality has, and the background image of the store has a kind of liminal quality to it – which can work in some situations – but doesn’t work here.
Let’s try another prompt:
A highly detailed and realistic cinematic still frame from a 1970s influenced horror film. A tall nude female with messy shoulder length platinum blonde hair, teardrop shaped breasts, wide hips, and a toned body. She is wearing a thigh holster with a beretta 92fs in it. She is standing in an abandoned underground laboratory that is dimly lit by the glow of computer monitors that display error messages. The laboratory has a retro futuristic design

So that hit our limit for Quality images.
Total Quality Images: 36.
Number of moderated Quality Images: 26:
So the Quality limit is probably somewhere around 50 to 60 images.
With moderated results, it’s easy to burn through them quickly – which is probably why the rate limits feel so constrictive.
But I think that’s pretty consistent with where we were in April too -It’s the reason why I haven’t really relied on the Quality setting for most of my images – when it works, it’s great, but it’s a little too erratic and scarce to be useful for much else.



Let’s see what Speed does with that prompt:


That’s probably about a 25% success rate, which is pretty low. But given that some of the results are producing full nudity, that also makes sense.








[SPEED IMAGE COUNT: 117]
So even though we triggered a rate limit with Quality, Speed still has generations left.
By my math, we’ve done 117 Speed generations, not counting moderated attempts.
And that’s pretty consistent with what the limits were before the Quality model was introduced, which I talked about here:
Now, how long did it take to blow through the Quality generations?
About an hour; I started at 11:39AM and the rate limit message appeared around 12:41 PM.


How long do those rate limits stay in place before they refresh?
I honestly don’t know.
At 8:52 PM the same day [approximately 7 hours later] I tried to generate quality images and this was the result:

So after about 7 hours I WAS able to generate Quality images again. I don’t know if that’s because there was a hard reset, or if it’s because there’s a rolling window based on the time at which generations occur.
So to recap:
I hadn’t heavily used Imagine for a few days – meaning that I hadn’t been hitting my image and video generation limits.
Basically starting fresh, I was able to generate 36 Quality images before I hit the rate limit. There were 26 moderated results which puts the total number of Quality generations at 62.
It took about an hour to hit that limit.
Based on my usage, I would say that’s where the rate limits have always been for Quality.
I was able to generate 117 Speed images without hitting any rate limits. Moderated resutls probably put the total at around 150.
A rate of 100 images per day was the normal rate prior to the introduction of the Quality model. That’s also pretty consistent with the email from xAI support that was posted by one Reddit user.
Quality appears to have a high rate of moderation – about 41% of all attempts were moderated.
Speed had a much higher rate of success with prompts: about 78% of all attempts were successful in generating an image.
So I don’t see a whole lot that’s different in terms of how many images I can generate now vs. how many images I could generate a month ago. But I know other people have had different experiences.
This is just a snapshot of what my image generations looked like today.
I haven’t tested out video limits – I’ll probably do that tomorrow and try to post about the results.
One thing I have noticed is that rate limit complaints tend to also complain about moderation.
So I want to be clear about something:
Moderation issues come down to things: bad images and bad prompting.
Some images are simply not going to produce the desired output if you are creating NSFW content. The odds of an uploaded image [that was not created in Grok] yielding an NSFW image or video are very low.
If your image generation prompts are getting moderated, you need to experiment and change the language you’re using to see what gets through.
If your video prompts are getting moderated, you have two choices: go extremely vague and see what slips through, or go for an extremely detailed prompt that does not leave Grok room to improvise something that will get moderated.
DO NOT SUBSCRIBE TO SUPERGROK HEAVY.
SuperGrok Heavy is not going to fix moderation issues.
Brute force should not be anybody’s primary strategy to generate adult content.
If you have the money to spend on SuperGrok Heavy, do yourself a favor and spend it on other models. For $100 to $300 a month, you can literally buy access to models that will make anything you want in the quantities that you want.
There’s a great list of Grok alternatives here:
And for anybody else that’s a SuperGrok subscriber that’s unhappy, you should absolutely cancel your subscription if you’re not happy with the service that you’re receiving. For $30 a month you can buy access to other models that may be less frustrating to use.
For now, I am just waiting to see what happens – but in terms of image generation, I don’t see anything that alarms me. Yet.






Leave a Reply