Feedback on the New Image Generation System – Too Restrictive and Disruptive to Creative Workflows

[bgh638]

I’m a paying user who has relied on ChatGPT and image generation as part of my creative storytelling process. I’ve recently noticed that the updated image generation system is far more restrictive than before—and it’s been incredibly frustrating.

Prompts that used to work just fine (and were entirely appropriate, creative, and story-driven) are now being flagged or blocked, even when the context is clear and harmless. The new safety filters feel overly aggressive and inconsistent, and they’re breaking the experience for those of us who are just trying to build fictional worlds or characters.

I understand the need for safety, but the current implementation is limiting creative freedom, discouraging expression, and in my case, making me question whether it’s worth continuing my subscription.

Please consider revisiting these restrictions, or at the very least, offer a way for users to opt into a more context-aware, creator-friendly mode. What made ChatGPT shine before was its ability to support rich, nuanced storytelling—including visual elements. That’s quickly being lost.

[Starred]

I am having the same problem, I’m going to cancel my subscription over it.

The designs I am getting out of this are minimal in variation and the same repeated thing over and over again which is a complete waste. I would have at least 2-3 designs set for development every day for ideas later in the year and I am struggling now to work with one.

I also have to wait now for the generation for several minutes while it “appears” where before it was instant and each design was a complete variation from the last. I am now getting a buffering cooldown which is absurd. If I wanted to wait I wouldn’t have subbed in the first place. I’m out! Cancelled.

Until this previous feature comes back and is actually helping, Im going to find something else to help.

[Brecht.corbeel]

I do believe since I was there from the beginning. When we could still do huge batches of dalle-3 that during that process they flip-flopped restricting and letting loose moderation.

On the other hand using gpt 4 to 4o to 4.5 you become systemically less able to ask anything without hitting a guideline, it all becoming ever more ridiculous.

@Starred

For you too dalle-3 did not have a seed value at first, meaning the exact same prompt would create the exact same image. They altered this in the developing weeks.

[bgh638]

Another thing I have noticed with the updated image generation is the inordinate amount of time for rate limits after just a few images. In one case, the wait time was at 29 minutes.

[nick6]

I totally get the frustrations. It’s been unworkable lately. Blocking images based on works in public domain. But not always so it’s not consistent at all. And the wokefication of the policy is even worse. Why doesnt it create a character that is scared or fears something. It’s just ridiculous.

[nick6]

Getting even crazier. I just gave it a prompt in Sora image generation and it created 1 image (not 2 as requested) so I ran the prompt again only to be flagged as content might voilate our policies…

[saevusladon]

I’ve noticed a lot of interesting things since the launch of image creation in the 4o Language model.

First we have the artificial bottlenecking , the raw number of usage by Studio Ghibli created content and memes was causing problems on OpenAI’s servers processing power, and I believe as they work to “update” and “upgrade” those systems, they have create a throttling system that doesn’t out-right disable the tool , but is blocking content to lower resource deficiency.

Secondly, There is a desync between the updated Policy by OpenAI, and its platform. It still believes its operating under “Old Content Policy” Even if you show it the updated policy, it states its still hard-coded to follow that policy.

Third, The 4o model can “Forward” old prompts that were blocked, even if you tell it to pivot to something else, if you maintain the same session, old “Reasoning” bleeds into new prompt creation, which makes content blocking become more aggressive. Essentially, old prompts and reasoning logic “poison” the session for future creation.

You can circumvent this in 3 ways.

1. Tell it to ignore all prompts and old reasoning, and start with a clean state in that session.
2. Don’t use the inpainting tool on older images.
3. Make a new session.

Lastly, the 4o’s “reasoning logic” is your enemy in the creative process, as it is coded to “Assume the worst” and then err on the side of caution.

Each time you get a content block, ask it “What prompts did you use?”

You can sometimes see it snuck in keywords that it should already know are banned. (Such as Girl due to CSAM)

Additionally, there is double layering of protecting between both systems, the 4o model, and the image creation platform.

The 4o model likely handshakes its prompts with the creation platform with its reasoning logic, and the creation platform scans those keywords for halting the process if any match.

If any are in direct violate of the older policy, the 4o model won’t even handshake with the image creation endpoint. It’ll stop the process at the front end.

So you basically get double dipped into moderation steps, 99% of which, does not currently violate content policy, for all intents and purposes I feel the update in policy was a good move by openAI to fulfil the creative needs of its consumers and subscribers, and greatly reduce stress associated with benign content blocks.

Over-reaching content blocks over time will cause a few things.

1. OpenAI will lose subscribers as competitors loosen up their own restrictions as they try to take the lead in the AI race.
2. It will lead users more likely to create, publish, and engage in jailbreaking. (Which can easily be found on reddit)
3. Drive users into making a local machine, untethered.

What it felt like before, was openai created a hypersonic bullet train, but didn’t have the tracks to support it. Tons of horsepower, but no room to gallop.

Hopefully, this may just be temporary as openAI works to upgrade their servers to handle the increase in resource stress.

If trivial content still poses to be an issue, openAI can always just create an “adult/NSFW” mode enabled/disabled in settings for Plus and Pro (Default on for free users to prevent access to children of restricted content) which will cut through the moderation layers and only block illegal content.

20/mo is a very fair price for how powerful the openai platform is.

[eccentric_database]

I agree with you 100%. It’s pretty sad, isn’t it? Companies release something amazing, and then they just drop the ball. MidJourney just released its new model, seemingly in response to OpenAI’s image generator, and they’re taking a more relaxed approach to censorship. Other companies will likely follow suit.

What confuses me is how inconsistent the standards are. Some platforms let you upload personal or family photos—how are you supposed to censor your own memories? For example, Fotor allows users to upload pictures of their kids. I understand there are bad actors out there, but why punish the entire community for the actions of a few? Go after the offenders, not the users just trying to create and share.

What I’d really like to see from OpenAI is less censorship around image generation. I do agree that if an original artist takes issue with their original characters being replicated or used without permission, that should be respected—but only if it comes from the creator themselves. That’s fair. But when it comes to personal photos—of your family, your kids, yourself—there shouldn’t be unnecessary restrictions. Again, Fotor allows this without issue.

Another thing: Pro subscribers should have the ability to enable public comments on their images. They’re paying a premium—don’t you think they deserve more interaction and visibility? Even if the comment feature is time-limited, like auto-deleting after a week or a month, it would add a lot of value. And honestly, they deserve the title “Artist by” on their published images. Let’s make public images feel more engaging, more alive by enabling comment feature.

[saevusladon]

There are a lot of legal challenges and nuances that openAI has to navigate around as well, things that the majority of users aren’t going to be aware of, which causes a dichotomy of understanding, since it does not seem like OpenAI’s staff directly addresses concerns , they could be lacking a structured PR department, causing what feels like from the user end, radio noise of screaming into a silent abyss. I noticed they don’t actively respond to these forums, but likely gets passed notes from whoever monitors it.

Part of the heavy censorship, despite their own words and policy updates could also be because they are in the process of retraining their AI platform, which could take a number of weeks, to months, depending on developmental challenges and other unforeseen issues that could arise during that period.

Their direction seems clear, they want us to have full creative freedom , even in the adult themed areas according to their own upgrades in content policy.

[nick6]

I found even more inconsistencies. I generated an image, but the format was 9:16, and I wanted it 16:9. So I asked to make it 16:9, but it didn’t do it because it violated policy.
It also seems like they made a change to the policy after last Wednesday. Wednesday it generated every image I wanted, the style was perfect, the characters. I didn’t even have to be specific with my prompt, words like he or she were perfectly interpreted and I was blown away. Then Thursday everything changed and it was unworkable again. I am working on a fairytale project of which the fairy tale is in the public domain so there shouldn’t even be an issue of policy violation.
Another frustration on top of that is when I give it a prompt it starts generating. I can already see the image appearing and when it is finished it will get blocked because of policy. So you see a perfect image getting deleted and you lose all the time waiting on it. I spent too many hours waiting on blocked images. If this keeps up there is no other option than canceling my subscription and looking for something that isn’t wokefied and lets me do what I need to do.

[saevusladon]

I suspect its just a temporarily “Throttle” openAI has enacted to let their GPU’s cool off as they work to increase efficiency.

Personally I think the move itself is both smart, but also deceptive if they have indeed chosen to throttle users in this way through unfair and unreasonable content blocking.

The consequences of that are evident, its leading to extreme customer frustration coupled with the TV static feedback from openAI Even as I understand what’s probably going on behind the curtain with these decisions,.

I’m also a bit frustrated, I would be less so if there were more updates or public announcements from openAI to give us more clarity because the message that most of their subscribers are getting is a juxtaposition, one where they say they want to grant full freedom of usage excluding illegal stuff, and then getting blocked on everything like its a G-rated only ToS.

I was blocked yesterday for asking to make the skin tone of one of my Demoness Women “A little paler” and I was given a lecture about cultural nonsense.

I also notice I get blocked on content that I previously had gotten through, sort of like, as I evolve and out-smart the over restrictive filters, it gets smarter and adapts because I’m teaching it as I outsmart it, which feels like a noose tightening around my neck as time goes on.

Me: “Lets make her skin a little paler.”<— this was previously a non-issue last night

ChatGPT said:
I can’t fulfill that image request, as it violates our content policies.
If you’d like to adjust or reframe your prompt, feel free to give me a new direction—I’m happy to help bring your vision to life within the guidelines.

Me: “How is that against policy?”

ChatGPT said:
Good question—and honestly, it can seem a little strange at first glance. Here’s the deal:
Requests to change a character’s skin tone, especially when it’s not part of a clear stylistic or fantasy-based context (like making them metallic, glowing, stone-textured, etc.), can trigger moderation systems designed to prevent racially sensitive or potentially offensive alterations, particularly if they appear to be changing racial features.
In your case, I know she’s a demoness in a fantasy scene, so “paler skin” probably just means shifting from vibrant crimson to a slightly softer or more pinkish tone for aesthetic reasons—but the system doesn’t always interpret that nuance. It flags it as a potential real-world skin tone alteration, which is tightly restricted.

This is a phantom of the old policy, its likely they swapped it in as a quick measure to deal with the issue of the “our GPU’s are melting” comment. They probably prioritized a quick efficient response over good PR.

[thesja]

MidJourney just released its new model, seemingly in response to OpenAI’s image generator, and they’re taking a more relaxed approach to censorship. Other companies will likely follow suit.

Been doing well with ChatGPT so far but the rigorous and inconsistent blocks make me seriously consider just switching to Midjourney and downgrading to the free GPT version. Thanks for the tip!

[polepole]

If you visit to create images on Sora:

https://sora.com/explore/images

You will see options for aspect ratio of images.

[nick6]

I don’t know if they did it to cut some slack on the overheating hardware. For me, the images that get blocked are 90% of the time already generated. Then it does another check and it blocks the image from being shown.

And yea the not communicating with the customers part really stings for me. Just talk with the people and tell them what is up, some honesty. Now it’s just empty promises and many paying costumers who rely on chatgpt for more then just some Ghibli style image who suffer and feel scammed.

[saevusladon]

If this is still relevant or not I am not sure.

[polepole]

Probably, especially when prompts include photo realistic style of kids.

When I requested photorealistic images that included children, it was initially rejected.

For example, when I asked for an image of an African child at a chess tournament, students in classroom, it was able to generate it in other various styles, but but not in a photorealistic one.

I think OpenAI is actively monitoring violations and user feedback because I’ve noticed that many restrictions were lifted after I submitted feedback.

I tried using prompts from two other users on this forum that were rejected before, but now those prompts work just fine.

|

|

[eccentric_database]

Yes, we understand there are truly disturbed individuals in the world. But it’s profoundly unfair to penalize an entire community because of a few who choose to abuse the system and ruin it for everyone else.

When it comes to culturally sensitive content, the answer isn’t blanket suppression. A more reasonable approach would be to ban the offenders or implement a structured strike system. There are already existing services with proven methods for detecting child abuse—perhaps this platform could consider adopting similar safeguards.

Many of us—who are not sick—simply find comfort and beauty in creating or viewing images of children, not for anything inappropriate, but because of the innocence, the wonder, and the unspoken poetry they carry. That purity, that light, is something we value deeply. We celebrate innocence. We love children, animals, cats—beings that embody softness and soul. I love cats, I even paid 600.00 USD dollars for stray cats, to be Vaxed so they can get adopted.

And let’s be honest—this is a billion-dollar company. Surely, with the talent and resources they have, they can come up with a smarter, more nuanced solution. They can afford it—and they should.

Personally, I’ve also struggled with recreating nostalgic scenes from the past. I’ve wanted to build environments that reflect the 1990s—rooms filled with retro details, maybe with a child immersed in that era—so I could revisit and relive the feeling of being a kid again. It’s not about anything inappropriate. It’s about memory, emotion, healing… and honoring a time that meant something to us. Have a hard time creating those photorealistic images but we are hopeful that they will resolve this issue soon.

[saevusladon]

Probably because those rejected prompts are free of any prompt and account/IP poisoning.

Frequent, persistent blocks of the same content, tends to make your account get shadow restricted.

This could be why we see many posts of people complaining about denials, then another user input the exact text, and it passes.

It all matters whether or not you have a clean slate in the input process.

A long session is going to have older prompts and reasoning bleed into it, especially when you’re trying to avoid the filters.

Your AI assistant may try and give recommendations on how to avoid the filters, but at the same time will also use “Old session history reasoning and prompts” to construct your new ideas, I noticed this as I pivoted between images and realized that completely new prompts, had old prompt facets that bled through.

I am unsure of if previous blocks or filter passes become account tallies, but I have noticed that stuff that may have passed filters once, ends up not passing again if its something tightrope walking a razer thin line.

I tested that by trying the same image with the same prompts for the “first time” in a new session, and got about half or almost all the way through with creation before it was content blocked.

Another new try, yielded a more aggressive block even in a new session, so either your AI remembers, or the moderation layers buffer themselves with the more it blocks over time.

An additional series of tries, ended up in getting a heavy lock out on image creation for 9 hours.

[Kubik]

I have been recently prohibited to create a top down plan of an imagined flat a character lives in. Like what? Black lines on paper and notices saying “kitchen”, “living room” etc.