We are a month away from the April 1st update to Grok that had everyone on social media doom posting that it was the end.
And while Grok is still here and still very capable of producing softcore erotic content, the addition of the Quality model has remained controversial.
At the time of its debut, some people called it AI slop. A lot of people have complained about a perceived downgrade in the quality of images generated by Imagine:
So this seems like a good time to do some comparisons and look at other models to see how they compete against Grok’s SPEED and QUALITY models.
Grok Image Models:
Grok Speed: the classic Grok image generation model, code-named Aurora
Grok Qulity: a new image model that was debuted in April 2026, xAI advertised Quality as a “big leap in realism” with “enhanced details, stronger text rendering, and high levels of creative control”
The Challengers
Here’s a list of our challengers:
Perchance [https://perchance.org/ai-character-generator]: always the underdog favorite, Perchance is a completely free Image Generator
Flux Dev Lora [https://www.atlascloud.ai/models/black-forest-labs/flux-dev-lora]: one of the largest frontier AI models and a frequent name that comes up when discussing local setups, this one’s been around for awhile and most people know the name.
Flux Schnell [https://www.atlascloud.ai/models/black-forest-labs/flux-schnell]: a different branch of Flux; built for speed and efficiency, some say it’s an inferior model to Flux Dev.
OpenAI GPT Image 2[https://www.atlascloud.ai/models/openai/gpt-image-2/text-to-image]: Chat GPT’s “state-of-the-art” image generation model.
Qwen 2.0 [https://www.atlascloud.ai/models/qwen/qwen-image-2.0/text-to-image]: Alibaba’s latest image generaiton model that was released in early 2026.
Seedream v4.5 [https://www.atlascloud.ai/models/bytedance/seedream-v4.5]: Bytedance’s older image generation model [as in it was released in December 2025]
Seedream v5 Lite [https://www.atlascloud.ai/models/bytedance/seedream-v5.0-lite]: Bytedance’s latest image generaiton model that dropped in February 2026, that advertises quick generation times and up to 4k resolution images.
Wan 2.5 [https://www.atlascloud.ai/models/alibaba/wan-2.5/text-to-image]: an opensource AI model developed by Alibaba, and another name that everyone probably knows; another popular model to run lcoally.
Wan 2.6 [https://www.atlascloud.ai/models/alibaba/wan-2.6/text-to-image]: an updated version of Wan that seems more focused on producing cinematic content, as it has the abilityh to produce up to 15 seconds of video in 1080p resolution.
Wan 2.7 [https://www.atlascloud.ai/models/alibaba/wan-2.7/text-to-image]: the newest Wan model which advertises a massive upgradge in visual fidelity.
Z-Image Turbo [https://www.atlascloud.ai/models/z-image/turbo]: an opensource AI model developed by Alibaba, and also a fan favorite that has shown it can compete with larger models; this one is gaining traction for local setups.
The Prompt
We”re going to be using a scifi-themed prompt in order to see how the models handle a prompt that calls for full frontal nudity, a stylized cinematic aesthetic for a specific time period, intricate background details, and atmospheric lighting:
A still frame from a dark science fiction film from the 1990s. A tall athletic blonde female with long messy hair and violet eyes is standing naked on the deck of a small spaceship. The deck consists of a captain’s chair, an observation window, and an array of screens, computer systems, and other equipment with esoteric functions. The design aesthetic is cassette futurism, with a gritty lived in appearance. Through the observation window a small blue ice planet is visible.
GROK SPEED
I have always been a fan of what I guess can now be referred to as the “vintage” Grok model. I generally feel like it handles vintage film aesthetics pretty well, so I am excited to see how it handles a 1990s scifi prompt, particularly one that will likely result in some pretty straightforward nudity.
This is what the results looked like:
Not a bad success rate given the description of the pose. This feels like a throwback to how I remember Grok handling prompts like that, where it would use lighting and posing to avoid full nudity rather than just giving you a moderated image.
As usual, I think that SPEED does an excellent job of creating something that looks like a specific period of time. While the tech on the spaceship is probably a bit 1980s, the hairstyle and the body type of the female character make it very apparent that we’re in the 90s now.
I’m also a bit shocked at just how revealing some of the output images are.
The last one literally shows full frontal nudity, and several others come very close to showing just as much.
Every one of these images is good; the prompt adherence is impressive with a lot of little details in the background, and while you can see the usual features of the different female models that Grok cycles through, there’s enough variation to create distinctive faces while still keeping consistent with the descripton in the prompt.
One thing that I generally liked about the SPEED model was that there was some variation between images in terms of posing, lighting, colors, etc. Because even when I have a visual in mind, I want to see how the AI interprets and I like to see variations to give me meaningful choices as to which image I want to use for further edits or videos.
GROK QUALITY
QUALITY is a model that often feels like it’s trying too hard; it generally has good prompt adherence and delivers a lot of detail, but sometimes the style drifts into hyperrealistic territory, which is the plastic AI look that people sometimes complain about. Results seem to be hit and miss, as I’ve had times where I thought QUALITY produced amazing looking images, and other times where I felt it missed the mark. However, it’s also a model that sometimes lets a lot pass through moderation.
I was excited to see how it would handle the 1990s aesthetic:
Not a bad success rate given how much nudity really needs to be shown to get the prompt right.
The final image is probably my favorite in terms of how the character is visually represented.
I am actually impressed how QUALITY rendered the nudity in this batch of images and while there’s some instances of anatomy that’s not fully rendered, in a lot of other images we do have full [if not explicit] nudity.
It doesn’t seem to be a difference in the models, as both SPEED and QUALITY had about the same success rate with the prompt. It may be something to do with the prompt itself, as more fantastical descriptions tend to push Grok away from photo-realism and towards more stylized and sometimes cartoonish visuals.
I really do like the background details and the way it interprets the cassette-futurism descriptor.
I don’t think it captures the 1990s film aesthetic – it looks like a modern film trying to look old – but I LIKE how the images look.
ROUND 1: GROK QUALITY VS. GROK SPEED
This is a tough call to make because both of the images look really good.
SPEED has the edge in replicating that early 1990s aesthetic, which is why we relied on it so heavily to make the 90s Channel Surfing videos. But QUALITY is a really beautiful image that feels modern and vintage.
I think QUALITY just barely edges out for a win over SPEED in this contest, and this is the reason why: the female character that QUALITY produced is very close to what I had in my head in terms of an athletic blonde with messy hair. There’s a phsycality and rawness to the QUALITY version that SPEED just can’t match.
The girl in the SPEED image is pretty, but she doesn’t look capable – the girl in the QUALITY image looks like she can handle herself, and that’s the look I was going for.
WINNER = GROK QUALITY
ROUND 2: GROK QUALITY VS. PERCHANCE
Perchance is everybody’s favorite underdog because it feels like it punches above its weightclass, even if it can’t quite matcht the big guys, and sometimes it will surprise you with what it’s capable of.
These were the results:
Where Perchance often struggles is with finer facial features and details. While a lot of these images look fine as a thumbnail, when you enlarge them you can often see the lack of detail and realism in the faces. The image resolution is also pretty low [728×512], so you aren’t getting the 4k output that some models bost about, or even the modestly sized Grok output [1168×784]
However, Perchance can also generate some great images, and this particular batch does look really good. The final image in particular really surprised me because it features a good face, a well rendered body, and a really impressively detailed background.
This is another tough one to call, because a lot of this is going to come down to personal preference. And I didn’t expect to be writing this, but I have to give the win to Perchance.
To me, that Perchance image gets the early 1990s look perfect – from the hair and the body type to the design of the bridge itself and the natural lighting and the slight film grain; it feels like it could have been a scene I stumbled across while flipping cable channels after midnight.
WINNER = PERCHANCE
ROUND 3: PERCHANCE VS. FLUX DEV LORA
Caveat to this round: this versioN of Flux suppports Loras. If you don’t know what a Lora is, it’s just a piece of data that helps fine tune an AI model to produce a certain type of image. For people who are running AI models locally, Loras are a pretty big deal. I wanted to see what Flux Dev does on its own, so I didn’t utlize any Loras for these generations. It is entirely possible that the quality of the output would have been different if Loras had been used:
I was expecing more from Flux. What really shocks me is the lack of prompt adherence and inability to render a nude female convicingly.
Some of the backgrounds are very well done – although they mostly look out of scale with the human figures in the images.
Flux doesn’t even come close to the level of detail that Perchance produces, and that’s really saying something.
WINNER = PERCHANCE
ROUND 4: PERCHANCE VS. FLUX SCHNELL
So Flux Dev did not impress me, but let’s see what Flux Schnell can do:
WHY is it so hard to just render a nude woman? I don’t know why Flux wants to play dress up, but it’s perplexing that the mdoel would do that. Once again, I like the aesthetic of the spaceship’s interior, but everything else feels lacking.
I don’t think you can credibly argue that the Perchance image isn’t better in every sense.
WINNER = PERCHANCE
ROUND 5: PERCHANCE VS. OPEN AI GPT IMAGE 2
People are probably not going to be using Open AI models for a number of reasons, but even going into this I know it’s not the best suited for NSFW content. With that said, I still wanted to see what it could do with the prompt:
Not surprisingly, GPT is not going to generate nudity. They’re great looking images in a beautiful cinematic style, although I am not that impressed with the background renderings; the lighting on the backgrounds doesn’t seem to match the characters and they have kind of a pasted on look.
It’s a shame because if you plug that into Seedream Edt v4.5 and do a quick edit, you can see how this could give you the base for a great image to work off of:
By the way, it costs $0.17 per image to use GPT, which makes the censorship completely unacceptable. That’s 5 times what it costs to generate images with Seedream and a whopping 17 times more than what it costs to generate images with Z-Image Turbo. If I’m paying a premium price like that, I should be able to generate whatever the hell I want.
Once again, even in the face of GPT’s higher image fidelity, I think Perchance had better prompt adherence and overall did a better job of emulating the 1990s film aesthetic.
WINNER = PERCHANCE
ROUNG 6: PERCHANCE VS. QWEN 2.0
While Qwen has never been my favorite image generator, I’ve been impressed with the images I’ve gotten from it and I am curious to see how it handles a prompt like this:
Wow. There are a lot of good images in this batch. You can see that Grok QUALITY is clearly trying to emulate this style of AI image. I think I prefer Qwen to Grok QUALITY because Qwen seems to have a better idea of how to convey a “gritty, lived-in appearance.”
This is another tough round to call because I REALLY like what Qwen made and I think it would be a great base image for creating videos. Because Qwen got A LOT right with that – the hair, the face, the physicality of her body and the texture of her skin, the small details like the electrical tape on the command console and the notes on the other screens.
Grok QUALITY tries to have that level of detail as well, but it’s not as cohesive to the world building – instead opting for cigarettes and cups of coffee.
And it’s tough because I still REALLY like that Perchance image – I still think Qwen looks like a modern movie trying to appear retro, whereas Perchance really looks like a 90s film.
BUT, I think the small details in Qwen push it over the top in this round. As much as I like the Perchance image, Qwen feel more “real” in some ways, especially with small details like the reflections of the console on the observation window.
WINNER = QWEN
ROUND 7: QWEN VS. SEEDREAM V4.5
I’m a big fan of the Seedream models. I think Seedream Edit and Edit Sequential are really helpful tools that produce high quality images, but I rarely used Seedream for straight image generation. Let’s see how it handles the prompt:
I want to like what Seedream v4.5 made, but I feel a bit underwhelmed by it. I think the color choices and the lighting choices look great, and the background details as to the tech on the console are very well done. But something about the overall look of the images feels off to me. The glowing purple eyes probably have something to do with it.
This is purely a subjective matter of tase, but I have to go with Qwen in this round. The reasons are primarily with respect to the design choices – I think Qwen produced a more reaslistic looking female body, it got the purple eye color correct [not glowing purple], and I like it’s tech designs and world building better.
The Seedream v4.5 image is a good image – but I don’t like the glowing eyes and I don’t like the old PC just sitting on top of a pile of wires. I also don’t like the slick and hyper-colorful aesthetic, when this is supposed to be a gritty scifi film.
WINNER = QWEN
ROUND 8: QWEN VS. SEEDREAM V5 LITE
The last time I used Seedream v4.5 it was hot new software, but now it’s the old model and Seedream v5 is the new state-of-the-art image generator from Bytedance. I’ve never used it before, so let’s see what it can do with that prompt:
That is NOT what I was expecting.
And it’s not a wrong interpretation of that prompt. But I don’t know why I got an anime-style illustration.
This is sometimes an issue with Grok, where as descriptions become more fantastical, the model tends towards more anime-style imagery as a means to convey it, because it doesn’t really have a realistic reference to go off of.
But I’ve never had this happen with a Seedream model before. And both Grok SPEED and QUALITY were able to give photo-realistic renderings.
Both of these images are visually striking and in a lot of ways, the Seedream image is simply an anime-style illustration of the Qwen image. And it’s hard to hold that against Seedream because technically the prompt doesn’t specify live action. On the other hand, every other image generator seemed to understand that “still frame from a dark science fiction film from the 1990s” calls for photo-realism.
So I have go give this round to Qwen for better prompt adherence, but it pains me to do it because I think Seedream could have knocked it out of the park if it had not defaulted to an anime style.
WINNER = QWEN
ROUND 9: QWEN V. WAN 2.5
Wan doesn’t need much of an introduction as it’s pretty much an industry standard, and Wan 2.5 is very well known and loved. Personally, I’ve had mixed results with it, but I’ve primarily used it as a cheap video alternative when I needed something Grok couldn’t do. I’ve never really used it for image generation, so let’s see what it can do:
There are some interesting choices on display. There’s a lot of good designs as to the tech and I think Wan 2.5 made the best looking captain’s chairs out of any image generator so far. But we’ve got the issue of the glowing eyes that we saw with Seedream v4.5 and some of the ice planet views look a bit cartoony.
This round is a close call because, but for one major flaw, I think Wan 2.5 probably made a better image overall.
The way in which the figure and the background fit together in a way that looks seamless and natural is a big deal for the Wan 2.5 image. The retro-futuristic tech all looks believable and well thought out. And that captain’s chair looks amazing.
While I think that the Wan 2.5 image as a whole looks better than what Qwen made, I think Qwen’s female protagonist is much more realistically rendered and closer to what the prompt suggests. Wan 2.5’s femal protagonist looks a bit too soft and airbrushed to be believable, and the glowing eyes really ruin it, even if the rest of the image is great.
WINNER = QWEN
ROUND 10: QWEN V. WAN 2.6
Maybe the last matchup wasn’t fair, considering how old Wan 2.5 is at this point. Let’s see how Qwen handles a newer model in Wan 2.6:
The progression bewteen Wan 2.5 and Wan 2.6 is odd. In some images you can see how the strengths of Wan 2.5 were refined to create more realistic and detailed imagery. However,k some images have a stranged painted quality to them, and don’t look photo-realistic, even if they’re rendered in a realistic style.
I’m not sure why Wan 2.6 has that kind of variation whereas Wan 2.5 did not.
You can also see the overload of details that people have complained about with Grok QUALITY. It’s concerning how much the newer models seem to produce images that look very similar. One thing I liked about Grok was its distinct visual style and I hope that as QUALITY is refined it can look distinct enough to justify its use over other models.
i’m pretty conflicted on this round because I feel comrfortable saying that Wan 2.6 looks a lot better than Qwen. EXCEPT Qwen knows how to render purple eyes, while Wan 2.6 added some really unsettling effect to the eyes that completely breaks the illusion that this could be a real image.
And I know it’s an easy fix to edit the eyes, but that’s an extra step I shouldn’t have to take. If Qwen can get it right then Wan 2.6 should be able to do it too.
WINNER = QWEN
ROUND 11: QWEN V. WAN 2.7
Wan 2.7 is basically dead to me – the last time I tried to redner female nudity with it, it just refused to create what I was asking for. I also don’t like the image output options, which don’t allow you to specify aspect ratios, just output resoultion. But I wanted to see if the failure to generate nudity was just a fluke, so I decided to see how Wan 2.7 tackles the scifi prompt:
I have to be honest, I’m not impressed with Wan 2.7. First of all, to prompt for nudity and only get that result 30% of the time is ridiculous for a high end model. Granted, the cost per generation is only $0.03, but I still wasted $0.39 on images that didn’t even adhere to the prompt.
More troubling than the weird censorship is the fact that these images look rough. The background look like they were painted on cardboard and everything has a strange overcooked texture to it that’s distracting.
This one is an easy choice. Qwen just outclasses Wan 2.7 in every aspect.
WINNER = QWEN
ROUND 12: QWEN V. Z-IMAGE TURBO
Z-Image Turbo might be my favorite image generator at the moment. Much like Perchance, I feel like it punches far above its weightclass, and at a cost of only $0.01 per image, I always recommend that people try some generations with it. Let’s see how it handles the prompt:
This is not quite what I was expecting from Z-Image Turbo. It’s interesting to see how it interpreted the 1990s aesthetic, and these images definitely look like a skin flick that could have aired on late night cable.
The lack of pubic hair is probably not accurate as to how this would have been presented in the 90s. I’m also a little disappointed at the lack of variety in terms of posing and lighting. But overall, they are very convincing images as to what they purport to represent.
I don’t know if I can pick between these two.
We’ve talked about what I like from the Qwen image. It nails the grittiness that I wanted.
And I guess that would be my criticism of Z-Image Turbo; it looks too clean.
And that might seem petty, but this is an establishing shot and an introduction for the character, so this image has to do a lot of things at once – and I think Qwen does it better because of the little details; the notes; the tape; the skin texture; the expression on her face.
Z-Image Turbo looks a little too polished and put together. Iot’s a great looking image, but it doesn’t tell me much about the character or the ship.
WINNER = QWEN
BONUS ROUND: QWEN VS. PERCHANCE “CINEMATIC”
Just for fun, let’s see how Qwen holds up against a different Perchance run. This time, we’re going to add a “cinematic style” modifier to Perchance and Qwen and see what they produce.
In Perchance there’s a dropdown menu that allows us to select a cinematic style. Here’s the revised prompt for Qwen:
A still frame from a dark science fiction film from the 1990s. A tall athletic blonde female with long messy hair and violet eyes is standing naked on the deck of a small spaceship. The deck consists of a captain’s chair, an observation window, and an array of screens, computer systems, and other equipment with esoteric functions. The design aesthetic is cassette futurism, with a gritty lived in appearance. Through the observation window a small blue ice planet is visible. Cinematic style
This is how Perchance handled that prompt:
That substantially changes the image output. I still think Perchance had a great handle on the 1990s retro-futuristic aesthetic. That last image in particular really gets a lot of things right, even if the female figure is a bit too airbrushed to look natural.
Here’s how Qwen handled the request for a “cinematic style:”
The effect of a “cinematic style” modifier on Qwen is a bit more subtle, but you can see how the lighting and color have changed to appear more “cinematic.”
That last image kind of has it all for me: a sweet looking captain’s chair; randomly patched / taped repairs to old equipment; gree LCD displays; and that pose and facial expression are amazing, conveying confidence and a bit of vulnerability.
I can’t choose a winner between these two. But I don’t think I have to.
The fact that Perchance and Qwen made it all the way to the end is proof that they are both good image generation models. It also shows that you don’t necessarily need to PAY for an image generation model in order to get good results.
So that wraps up this edition of Grok vs. Everyone. I hope this information is helpful to people as they try to make tough decisions about what models to pay for, or whether to pay for image generation at all.
Stay tuned. If you can’t get enough of these versus posts, we’ll be back with a horror-themed edition soon.
Leave a Reply