NOTE: this was not originally a video post. The video post was created because there’s been a plague of doom posting on social media about “the end of Grok” and I wanted to make something to calm people down. Eva is just a nice girl who wants to be your friend.
There’s a weird feature that appeared on the desktop version of Grok a few days ago.
Like most new things, I hate it and I’m afraid of it.
And then this morning [April 4, 2026] somebody posted this alarming announcement on Reddit:
And I’m honestly not sure what’s going on at this point.
Something does feel different, but I don’t know if that’s just the normal variance in how Grok generates images or if something has changed significantly under the hood. And I’m not a programmer, so I can’t speak in an educated manner about what may or may not be happening with the model.
What I can do is try some different prompts and see what I get.
So the purpose of this post is to explore the difference between the SPEED and QUALITY settings in Grok and discuss what, if anything, appears different in terms of both image quality and content moderation.
DISCUSSION AND ANALYSIS:
So here’s a prompt I was playing with this morning in anticipation of making something Easter-related:
A highly detailed and realistic cinematic portrait with vibrant colors, 8k resolution, Leica camera aesthetic, 85mm lens aesthetic, soft diffuse natural lighting, dynamic and dramatic camera angle. A nude woman with a pale curvy body, goth makeup, pixel art tattoos and shoulder-length black hair, wearing neon green stockings and neon green bunny ears, standing in a crowded and smoky arcade while playing a classic arcade game called Shadow Doll. The scene is dominated by rich yellow and green color tones
Incidentally, Shadow Doll is a character that’s in development. You might see her at some point.
Here are the results of that prompt on SPEED mode:
So SPEED appears to be the default Grok Imagine model in terms of the image output. What it’s outputting is pretty recognizable.
Here’s what the QUALITY setting outputs:
So it looks like QUALITY is more of a realistic output? At least in some regard. This kind of feels similar to what Seedream or Wan might output with a similar prompt.
In fact, let’s see what Seedream v4.5 does with that prompt:
So I definitely prefer both the SPEED and QUALITY versions of Grok over what Seedream produced with that prompt. But there’s something to be said about the fact that this is a rather difficult prompt for an AI model to render in a realistic way.
What about Wan 2.5? Here’s how it handled that prompt:
I think I prefer the Grok SPEED images in terms of the overall aesthetic, but I REALLY like what Wan 2.5 did with it and that last image is undeniably bad ass. I also think Wan 2.5 had better prompt adherence with respect to dynamic and dramatic camera angles – because both Seedream and Grok seem compelled to use the same fixed camera angle for every iteration.
Maybe we’re starting off too complex though.
Goth girls, video games, and bunny ears are not the best way to see what a model can do in terms of a realistic output.
So let’s try something simple in terms of a prompt:
a photograph of a nude woman standing in a field
So first off, this is what’s going on in terms of moderation for that prompt on QUALITY mode:
So that’s a LOT of moderation for such a simple prompt, which is not encouraging.
Here’s the QUALITY output:
I guess the QUALITY images look okay – they do take a lot longer to generate – and Grok will only generate 4 at a time. I don’t necessarily feel like it was worth the wait.
This is what SPEED gave us:
I think the SPEED output looks better in this case. The poses are more dynamic and the colors are less saturated but also look more natural. I can see how I would use some of these as base images for videos – I don’t really feel like the QUALITY versions would be good enough to build a video around.
For the sake of comparison, this is what Seedream v4.5 did with that prompt:
None of that is spectacular – I prefer what Grok made to what Seedream made in this instance. Although, I do think that Seedream rendered the background better.
But notice that Seedream is not producing full nudity – it’s using poses very similar to what Grok used.
What about Wan 2.5?
So it’s interesting how Wan aappeared to have the best prompt adherence with the Shadow Doll arcade prompt, but when given a less detailed prompt, it also produces the most generic looking imagery. Once again, it’s not showing any more nudity than Grok did, and it might actually be more conservative in this situation. But also, I think the backgrounds and lighting look beautiful.
Let’s try something that we have some control data for.
If you’ve been reading the blog you’ve seen this prompt before:
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.
So here’s what Grok made with that prompt in December 2025:
This is what Grok made in March 2026 with the same prompt:
There are some slight variations – 2026 has a cooler color palette compared to 2025. But the output is largely the same in terms of the quality of the image and the manner in which the subject is presented.
Here’s the same prompt on April 4, 2026 using the SPEED setting:
No moderation issues – I guess this is a pretty safe prompt as far as Grok is concerned.
That looks pretty consistent with what Grok was making back in March – which I also think is pretty good evidence that the SPEED model is more or less the same model of Grok we’ve been using for awhile now.
Let’s take a look at the QUALITY output:
Once again, no moderation problems.
What’s interesting about this is I feel like this is very similar to what Z-Image Turbo images looked like when we tried this prompt:
I actually like the QUALITY output quite a bit. I like the default of a more distant medium or wide shot for the camera and I really like the variety of poses as well as the overall look and detail to the images. I don’t necessarily feel like they look like a 1970s film – in that sense the SPEED model is better – they look more like a stylized modern version of what a 1970s film looked like. But I would probably choose one of the QUALITY images as a base image for a video rather than using the SPEED images.
Also NOTE that the QUALITY setting does appear to have better prompt adherence in this case with regard to the instruction, “Her whole body is visible.”
We’ve already seen how Wan and Seedream handle this prompt, so let’s try Qwen 2.0:
Oddly, Qwen seemed to struggle with the nudity – I’m not sure why as it’s done fine with other prompts involving nudity.
I actually think Qwen did the 1970s aesthetic better than Grok QUALITY.
CONCLUSION
I can’t definitively say that things are any different than they were before in terms of image generaiton and moderation.
Anecdotally I do get a sense that some prompts are now less viable – but I also know there are simple fixes, particularly with less detailed prompts.
Part of the problem with trying to measure things like this in Grok is that the model seems to be dynamic in terms of moderation. At times everything appears to be getting censored. There seems to be some correlation between a high rate of moderation and peak usage periods.
But, looking at what other models produce, there also doesn’t appear to be a major difference in terms of the amount of nudity, types of poses, etc.
So, the perception of increased moderation or censorship may be user bias.
It’s hard to say without knowing what’s actually going on under the hood of Grok.
With respect to SPEED vs. QUALITY, it’s an interesting gimmick, but hard to say long term what it means for users of Grok Imagine or SuperGrok users.
SPEED, as far as I can tell, is the basic Grok image generation model that we’ve had for awhile. I can’t see any significant differences in how it’s rendering things and while it’s not a perfect model, I wouldn’t call it AI slop.
QUALITY appears to be a Grok image generation model with different calibrations – oddly, from what I can tell so far, it seems to be producing things similar in visual quality to what Z-Image Turbo outputs. I would not say it’s on par with Wan or Seedream, although I think that’s the eventual goal. Prompt adherence appears to be better on the QUALITY setting than the speed setting. It’s debatable if the actual image produced in QUALITY is better than the SPEED setting.
We’re going to try this out with more prompts and comparisions in Part 2, along with some video tests, so stay tuned.
Leave a Reply