Right? Came to the comments specifically for this, but am confused by people's responses. With prompt adherence this bad, is it worth the 2 cents you spent on it? I don't see how it's even useful for deciding if you want to use the ultra version, or for anything else really.... Maybe if you want to redo it in Photoshop? But at that point, breaking out the old Wacom tablet and making a composite image would probably be just as time intensive, but with much higher image quality (and none of the tale tell signs of AIgen)
> Imagen 4 Ultra: When your creative vision demands the highest level of detail and strict adherence to your prompts, Imagen 4 Ultra delivers highly-aligned results.
It seems that you may need the "Ultra" version if you want strict prompt adherence.
It's an interesting strategy. Personally, I notice that most of the times I actually don't need strict prompt adherence for image generation. If it looks nice, I'll accept it. If it doesn't, I'll click generate again. For creativity task, following the prompt too strictly might not be the outcome the users want.
I've found this is an interesting balance with Copilot specifically. Like, on the one hand I'm glad it aims for the bare minimum and doesn't try to refactor my whole codebase on every shot... at the same time, there's certain obvious things where I wish it was able to think a bit bigger picture, or even engage me interactively, like "hey, I can do a self-contained implementation here, but it's a bit gross; it looks like adding dependency X to the project keeps this a one liner— which way should it go?"
I’ve had good experience with iterative prompting when generating images with Gemini (idk which model — it’s whatever we get with our enterprise subscription at work, presumably the latest.) It’s noticeably better than ChatGPT at incorporating its previous image attempt into my instructions to generate the next iteration.
Midjourney scores the absolute lowest in terms of prompt adherence against any of the other SOTA models (Kontext, Imagen, gpt-image-1, etc). At this point, its biggest feature is probably as an "exploratory tool" for visualizations by cranking up the chaos and weirdness parameters.
To the left of the "detailed spaceship" I think I see a distortion pattern reminiscent of a cloaked Klingon bird of prey moving to the right. Or I'm just hallucinating patterns in nebular noise.
In the little experimentation I did with AI image generation, it seems more a game of trying multiple times until you get something that actually looks right, so I wonder how many attempts they did.
Clicking on "Read the documentation" leads to a page that documents nothing about the latest Imagen models and only provides examples using Gemini 2.0 Flash.
I have found Imagein to be a good general purpose editor and we use it to clean up bitmaps, and adjust black points and white points and curves on greyscale, so it is good for preparing B&W greyscale photographs for print to compensate for dot gain in halftone screens on laser printers. Its 'color separation' capability is rudimentary/first draft though and is ridiculously close to inverse RGB rather than CMYK. For good color seps we use Photoshop so I can control undercolor removal.
They're probably talking about the original Imagen printing product line from the 1980's. I thought I might be the only one to remember them in this thread, so I did a search for printer and found the GP comment.
I guess it's kinda nicely genuine that the "four panel comic strip" has some errors in it (misunderstanding caption + cat high-fiving itself in the bonus fifth panel)
1. The text ”Imagen 4 is now generally available!” is still spoken, not a caption.
2. ”low latency” -> ”low-laten”
(3. Has that ugly gpt-image-1 trademark yellow filter requiring work in post to avoid.)
I didn’t bring up the ”retro comic look” thing. I certainly think it’s an issue with Imagen 4’s version. It doesn’t look very old school at all. But I can’t judge the OpenAI one either on that, I’m no comic book expert, so I just skipped that one.
The pervasive yellow tinge indicates that that is almost assuredly `gpt-image-1` - OpenAI's flagship model and (aesthetics aside) the highest scoring model in terms of strict prompt adherence that I've seen.
With images and video, it's less clear exactly what they're doing, but it's watermarking on the pixel leve. From one of their blog posts:
Videos are composed of individual frames or still images. So we developed a watermarking technique inspired by our SynthID for image tool. This technique embeds a watermark directly into the pixels of every video frame, making it imperceptible to the human eye, but detectable for identification.
> I didn’t bring up the ”retro comic look” thing. (…) I’m no comic book expert, so I just skipped that one.
I’m no Scott McCloud, but the OpenAI version definitely does a better job with the retro style. The yellow filter you criticised actually helps to sell the illusion. The Imagen version utterly fails in the retro area, that style is very much modern.
But there are other important flaws in the OpenAI version. The fourth panel has a different cat (the head shape and stripes are wrong) and it bleeds into the previous panel. Technically that could be a stylistic choice, except that the floor/table is inconsistent, making it clear it was a mistake.
I've updated my GenAI Comparison site to include Imagen4 Ultra, so now we have four Google related generative models (Gemini Flash, Imagen3, Imagen4, and Imagen4 Ultra).
Despite claims that Ultra supports improved strict prompt adherence, we saw no evidence that it scored any better than Imagen 4 and in some cases seemed to ignore the prompt altogether (see the "Not the Bees" comic). In many cases, it also seemed much less steerable than Imagen3 requiring many of the prompts to be rewritten.
There's some speculation it's Gemini 3's multi-modal output, and other speculation that it's an OpenAI model. Hard to definitively since these models tend to hallucinate when interrogated.
Other than LMArena and a website I can't verify is authentic, it's hard for me to run tests on this new model but I have serious doubts that it'll pass my more difficult prompts such drawing a valid 2d maze with clearly marked exit and entrance.
gpt-image-1 is in a class all of its own with regards to prompt adherence in the "text to image" category.
Once it hits GA I'll put it through its paces and add it to the site!
I tested it with generating a man holding a Penrose triangle made of wood. While gpt-image-1 succeeded, nano-banana failed. The aesthetics of nano-banana did look much better though. I would guess that it is a diffusion model, based on the fact that it adds irrelevant but pretty background details, which gpt-image-1 tends to avoid.
I tried the following prompt and other than producing a four panel comic that was black and white it completely ignored every other instruction. This was with 4 ultra. Maybe someone else will have better luck but the failure seemed stable.
'''
A four panel comic strip. Simple black on white. Stick figures for characters. In the first panel there is a stick figure man and a stick figure bird eating bird seed at his feet. He is slightly hunched over to show he is looking at the bird. In the second panel. He is more hunched over looking more closely at the bird. In the third panel he is even more hunched over practically with his head to the bird, he is crouched down, knees bent, hands on thighs. In the upper left of the third panel the tip of an enormous beak can be seen, but it's only a few lines so could be anything. In the final panel the beak has gobbled up the man and his arms and legs are flailing outside of the beak while the small bird continues to eat birdseed on the ground.
'''
In the couple of prompts I gave it, it's better than the last version but I feel that Google is sacrificing quality for the sake of speed. While it's a lot faster, the output is not as good as OpenAI.
Meanwhile Veo3 is far better than the OpenAI's equivalent. I assume speed is not a priority there; both take their time.
I'm glad they can't. The reason large cooperations tend to suck is because some bored management guy cares about typos and invents a process for getting your headlines approved by some other dude who is just as bored and useless.
I am currently building an AI product which relies on Imagen 3 to generate a lot of photorealistic, cinematic or HDR images. I tried Imagen 4 during preview, but results were too "cartoonish". Did anyone else have the same experience?
Yeah me too, I think 3 does a much better job for photos or even just images that look like realistic renders. I use 3 for generating grids of age-progressed portraits for a game and it does a better job at sticking to the prompt. 4 also seems to spit out ones that have that really smooth look that makes it really obvious it’s AI.
I hate that they always announce their image models months before they make them available. They should just announce them later. OpenAI does this much better, with a few days delay at most.
>The model may output text only. Try asking for image outputs explicitly (e.g. "generate an image", "provide images as you go along", "update the image").
>The model may stop generating partway through. Try again or try a different prompt.
The comments here are priceless. In less than five years time we have gone from "That's impossible" to "Meh, it doesn't solve P=NP if prompted.".
For those commenting in the latter category, it might be worthwhile to read a bit about the underlying technology and share your insights on why it does not deliver.
The problem with 2025 is I have seen thousands of better examples than that landscape. The reflections in the lake are complete trash.
Then I think of Veo 3 that is just incredible. So no, it is not impressive if a still from the video model is vastly better than the static image generator from the same company.
I find it especially annoying because I can't think of another company this would happen at. It is just so Google.
this is false and the two things are not correlated.
if you followed news during the GAN cycle you could extrapolate that deep NN could do this type of things. it is really cool that this things happened so fast, but we are talking about companies that have the money to deploy thousands of cars around the globe to collect data, so they absolutely know how to gather data
The webcomics is awful. It feels off, the characters look very fake, unsettling in the way they communicate. The prompt is shown bellow the image, but for me the result looks closer to a prompt "Create lifeless characters reciting marketing slop. They must fake an over exaggerated excitement but it should be clear they don't believe in what they're saying and have no souls".
Also, the prompt specifically ask "Panel 4 should show the cat and dog high-fiving" but the cat is high-fiving ... the cat. Personally I find this hallucinated plot twist good, it makes the ending a bit better. Although technically this is demonstrating a failure of the tool to follow the instructions from the prompt. Interesting choice of example for an official announcement.
It’s more than visual preferences — his image actually adheres to the specified requirements. it hasn’t been shown that Imagen can do that, which might be a showstopper for many people, regardless of aesthetics.
An image has to be much worse than that to fail to save you 6 seconds.
That said, this is their own chosen example of what it can do, so I'd have to assume it is much worse than that on average.
It seems that you may need the "Ultra" version if you want strict prompt adherence.
It's an interesting strategy. Personally, I notice that most of the times I actually don't need strict prompt adherence for image generation. If it looks nice, I'll accept it. If it doesn't, I'll click generate again. For creativity task, following the prompt too strictly might not be the outcome the users want.
https://i.imgur.com/kSuqCYg.jpeg
https://tug.org/TUGboat/tb02-2/tb03imagen.pdf
1. Not seen browsing ”ai.dev”.
2. The text ”Imagen 4 is now generally available!” is spoken, not a comic caption.
3. Invalid second panel.
4. Hallucinates ”Meet Imagen 4 fast!”
5. Hallucinates ”It offers low..” etc. (this is the second part of a single sentence said by the cat)
6. Hallucinates ”You can export images in 2K!” (this sentence is not asked for)
7. Doesn’t have the cat and the dog in the fourth panel.
—
Here’s the gpt-image-1 counterpart with the issues I could find:
https://chatgpt.com/share/689f7e4b-01e4-8011-8997-0f37edf8c2...
1. The text ”Imagen 4 is now generally available!” is still spoken, not a caption.
2. ”low latency” -> ”low-laten”
(3. Has that ugly gpt-image-1 trademark yellow filter requiring work in post to avoid.)
I didn’t bring up the ”retro comic look” thing. I certainly think it’s an issue with Imagen 4’s version. It doesn’t look very old school at all. But I can’t judge the OpenAI one either on that, I’m no comic book expert, so I just skipped that one.
https://i.imgur.com/kSuqCYg.jpeg
https://genai-showdown.specr.net
Repo: https://github.com/google-deepmind/synthid-text
Paper: https://www.nature.com/articles/s41586-024-08025-4
With images and video, it's less clear exactly what they're doing, but it's watermarking on the pixel leve. From one of their blog posts:
https://deepmind.google/discover/blog/watermarking-ai-genera...Elevenlab's audio watermarking is trivial to shake off with compression, but google claims that synthid is resilient to such manipulation.
I’m no Scott McCloud, but the OpenAI version definitely does a better job with the retro style. The yellow filter you criticised actually helps to sell the illusion. The Imagen version utterly fails in the retro area, that style is very much modern.
But there are other important flaws in the OpenAI version. The fourth panel has a different cat (the head shape and stripes are wrong) and it bleeds into the previous panel. Technically that could be a stylistic choice, except that the floor/table is inconsistent, making it clear it was a mistake.
Despite claims that Ultra supports improved strict prompt adherence, we saw no evidence that it scored any better than Imagen 4 and in some cases seemed to ignore the prompt altogether (see the "Not the Bees" comic). In many cases, it also seemed much less steerable than Imagen3 requiring many of the prompts to be rewritten.
https://genai-showdown.specr.net?models=IMAGEN_3,IMAGEN_4,IM...
There's some speculation it's Gemini 3's multi-modal output, and other speculation that it's an OpenAI model. Hard to definitively since these models tend to hallucinate when interrogated.
gpt-image-1 is in a class all of its own with regards to prompt adherence in the "text to image" category.
Once it hits GA I'll put it through its paces and add it to the site!
Muphry's Law strikes again.
Indeed.
''' A four panel comic strip. Simple black on white. Stick figures for characters. In the first panel there is a stick figure man and a stick figure bird eating bird seed at his feet. He is slightly hunched over to show he is looking at the bird. In the second panel. He is more hunched over looking more closely at the bird. In the third panel he is even more hunched over practically with his head to the bird, he is crouched down, knees bent, hands on thighs. In the upper left of the third panel the tip of an enormous beak can be seen, but it's only a few lines so could be anything. In the final panel the beak has gobbled up the man and his arms and legs are flailing outside of the beak while the small bird continues to eat birdseed on the ground. '''
Meanwhile Veo3 is far better than the OpenAI's equivalent. I assume speed is not a priority there; both take their time.
One of the biggest corporations in the world and they can't re-read before posting a typo in the title.
Heads be shakin
It's a typo, it doesn't matter.
>The model may output text only. Try asking for image outputs explicitly (e.g. "generate an image", "provide images as you go along", "update the image").
>The model may stop generating partway through. Try again or try a different prompt.
Seriously?
For those commenting in the latter category, it might be worthwhile to read a bit about the underlying technology and share your insights on why it does not deliver.
The problem with 2025 is I have seen thousands of better examples than that landscape. The reflections in the lake are complete trash.
Then I think of Veo 3 that is just incredible. So no, it is not impressive if a still from the video model is vastly better than the static image generator from the same company.
I find it especially annoying because I can't think of another company this would happen at. It is just so Google.
if you followed news during the GAN cycle you could extrapolate that deep NN could do this type of things. it is really cool that this things happened so fast, but we are talking about companies that have the money to deploy thousands of cars around the globe to collect data, so they absolutely know how to gather data
Also, the prompt specifically ask "Panel 4 should show the cat and dog high-fiving" but the cat is high-fiving ... the cat. Personally I find this hallucinated plot twist good, it makes the ending a bit better. Although technically this is demonstrating a failure of the tool to follow the instructions from the prompt. Interesting choice of example for an official announcement.
https://i.imgur.com/kSuqCYg.jpeg
Does the world need yet another AI slop generator?