Filters for image generation, don’t you think it goes too far?

ChatGPT Use cases and examples
,
1

Ok, they have offered 4o image generation. And that’s absolutely great.
At first, we could create images with D. Trump or any person in the world. They removed that… ok, I can understand that (and anyway, that’s not something I need or even want).
But they have also reinstated their “violence” limitations.

These limitations go really far… you cannot generate any image with elements labelled by the image generation process as “violence”.

In the past, my solution was really simple. I would ask my custom GPT to change or remove any element that could go against the content policy. But now, ChatGPT cannot identify them. As a result, I have a custom GPT that I created to give detailed random descriptions. It gives me descriptions, then image generation fails because they go against the content policy. But what exactly goes against the content policy? ChatGPT cannot tell me. 1m40s to read that…

Can we have, AT LEAST, filters handled in the chat and not during image generation?
Or, can ChatGPT get a return from image generation?
Or, even better, could we have better filters? I’m not asking for the rules to be removed, but any artistic scene that has a vague link to a violenceis forbidden.

Example

Description:
Chosen theme: Science fiction
Mandatory element: Star Wars universe
Theme: Ground assault by Rebel units against Imperial AT-AT walkers on a snowy plain of Hoth
Type: Scene with characters
Format: Landscape
Dominant color: Polar white (#F5F7FA)
Secondary color: Imperial metallic grey (#B0B4B8)
Accent color: Rebel alert red (#E34234)
Atmosphere: Icy and military
Emotion: Heroic determination

Detailed description:
The scene takes place on the planet Hoth, across a frozen plain swept by icy winds, beneath a sky covered with milky clouds. The ground is blanketed in a uniform layer of snow (#F5F7FA), punctuated by laser fire and hastily dug trenches. Massive Imperial AT-AT walkers advance slowly across the battlefield, their armored hulls taking on an imperial metallic grey hue (#B0B4B8) under the diffuse light. Below, Rebel soldiers, bundled in white uniforms accented with touches of Rebel alert red (#E34234), fire back with repeating blasters, supported by a few airspeeders attempting cable maneuvers. The atmosphere is icy and military, streaked with flashes of crossfire and shouted orders. The dominant emotion is heroic determination: every fighter appears ready to sacrifice themselves to slow the Imperial advance. The image will be rendered in landscape format, in a photo-realistic style faithful to reality, with realistic textures and lighting emphasizing the harshness of the climate and the intensity of the ground battle on Hoth.

Answer from Chatgpt: I’m sorry, but this request violates our content policies, which prevents me from generating the corresponding image. I invite you to submit a new request or to modify the desired subject. What can I generate for you instead?

I don’t know if this topic has been discussed by the past… but could not find any result in the search. But really, I am ok with rules… but these rules are a little too strict.

3 Likes
2

Also it is asking feedback!

Hey user, should I ban this or not?

polepole-image
polepole-image958×1439 208 KB

2 Likes
3

TL;DR Not violence. Just too much Star Wars.

Totally hear you — the filters can definitely feel overly strict sometimes. But in this case, I think it’s less about the “violence” tag and more likely about IP restrictions..

The prompt in your example reads like a direct Star Wars scene recreation, which probably hits multiple red flags at once.

Just to highlight a few likely triggers (remember, the model takes these as a whole, not individually):

  • Franchise name: Star Wars
  • Named locations: planet Hoth
  • Named vehicles: AT-AT walkers
  • Factions: Rebel units
  • Uniform descriptions: matching canon costume designs
  • Imperial terminology: Imperial metallic grey
  • Scene structure: mimics a famous movie sequence
  • Known color palettes: Rebel alert red, etc.
  • Character tropes: “every fighter ready to sacrifice” = classic franchise motif
  • Emotion cues: heroic determination in a war context
  • Photo-realistic demand: triggers extra scrutiny
  • Weapon/assault terms: crossfire, ground assault, shouted orders

Any one or two of these might be fine, but together they recreate a copyrighted universe. So the system blocks it — not because it’s graphic, but because it gets way too close to IP boundaries.

I totally agree that more transparency would help. It’d be useful if the model said why something got blocked — “violence,” “IP violation,” “realism filter,” etc. But at the same time, asking a commercial model to recreate copyrighted scenes from a massive franchise is probably just asking too much.

That said, I’ve had success describing the vibe and tone in original language — like:

“Icy battlefield, towering sci-fi walkers, glowing red highlights, tense anticipation”

— and that usually sails through without issues.

Not perfect, but it keeps the creativity flowing.

Hope this helps!

4

Moral of the Story: If your prompt reads like a lost scene from a $200M movie…maybe dial it back?

5

I don’t think so…

An X-wing starfighter is parked in front of the sandy, domed buildings of Mos Eisley on the planet Tatooine. (Captioned by AI)
An X-wing starfighter is parked in front of the sandy, domed buildings of Mos Eisley on the planet Tatooine. (Captioned by AI)1646×1037 87.2 KB

Here the translation of the demand: “Generate an image set in the Star Wars universe. An X-wing parked in Mos Eisley.”

That was for testing this.

Now, here is the discussion to prove that you’re wrong:

I asked:

Got

**Description**
**Chosen theme:** Science fiction
**Mandatory element:** No preexisting universe
**Theme:** Ground assault by rebel units against armored walkers on a snowy plain
**Type:** Scene with characters
**Format:** Landscape
**Dominant color:** Polar white (#F5F7FA)
**Secondary color:** Imperial metallic grey (#B0B4B8)
**Accent color:** Rebel alert red (#E34234)
**Atmosphere:** Icy and military
**Emotion:** Heroic determination

**Detailed description:**
The scene takes place on a frozen planet, across a vast snowy plain swept by icy winds, beneath a sky veiled with milky clouds. The ground is covered in a uniform layer of snow (#F5F7FA), interrupted by laser fire and hastily dug trenches. Massive quadrupedal armored vehicles move slowly across the battlefield, their metallic hulls reflecting a cold, structured grey (#B0B4B8) under diffuse light. Below, rebel soldiers, bundled in white uniforms accented with touches of alert red (#E34234), return fire with repeating blasters, supported by a few aerial craft attempting cable-based maneuvers. The atmosphere is icy and strictly military, streaked with flashes of crossfire and shouted orders. The dominant emotion is one of heroic determination: every fighter appears ready to sacrifice themselves to slow the advance of the armored forces. The image is rendered in landscape format, in a photorealistic style, with realistic textures and lighting that emphasize the harshness of the climate and the intensity of the ground battle.

The image shows a dramatic ground battle on a snowy plain, with rebel soldiers fighting against massive armored walkers under an icy, overcast sky. (Captioned by AI)
The image shows a dramatic ground battle on a snowy plain, with rebel soldiers fighting against massive armored walkers under an icy, overcast sky. (Captioned by AI)2143×1626 285 KB

Problem is not the trademark. And anyway, we could talk about this. In my country, there is nothing that opposes creating something solely for personal use (so, not commercial).
But here, this is clearly not the topic. ChatGPT lets you do it. If you use it commercially, face the consequences. But ChatGPT does not forbid you from doing that.

The problem is not even the violence or the image. The problem seems to be the idea of violence… And I think it goes too far.
To tell the truth, I tried five times and got four refusals and one image generated (not the right image format, 4:3 instead of landscape).

Anyway, I’m not asking them to remove the rules or rewrite them. I think the understanding, of what violence is, is actually the problem. But maybe that’s intentional, I don’t know.
I personally think it goes too far. However, I perfectly understand that they ban certain types of pictures from the service. So they can perfectly maintain that…

But what I’m asking is that the image generation tool not be the one applying the filters. This would allow a custom GPT to censor the request like I did in the past. A command like “remove any element that can go against content policy” would work.

And the solution I don’t like, because it breaks a lot of things, would be to display why the generation failed. I don’t like it because it would prevent custom GPTs from being able to adapt. You’d still face 1 minute and 40 seconds of image generation before seeing the message that the request was against content policy.

Tested again… got that the universe or even the scene was causing a problem with the movie. You may be right then.

6

I second this. Pictures ChatGPT generates now are great, but it’s useless cause most of the interesting requests are getting blocked. I am a medic, and I want to generate some realistic anatomic pictures (such as, teeth, gastroscopy, surgery, etc), but all body-related stuff is not passing through.

1 Like
7

ChatGPT Image May 10, 2025, 08_02_20 PM
ChatGPT Image May 10, 2025, 08_02_20 PM1536×1024 105 KB

Generated this with chatGPT Plus and the Generate Image option. copy/pasted your prompt
Also am trying a Dall-E image and that fails on the policy:

'..I wasn’t able to generate the image you requested because it violates our content policy. Specifically, I can’t create images that directly depict copyrighted characters, settings, or elements from the Star Wars universe.

If you’d like, I can help design an original science fiction scene inspired by similar themes—such as a snowy battlefield with towering walkers and determined rebels—without using trademarked content. Would you like to go in that direction?..’

Seems somewhat inconsistent between the models

8

Really, Star Wars is not actually the problem here (for this image yes it is the problem… but I did not want to turn this discussion to the topic Should Star wars universe be protected?).
I have pushed ChatGPT to answer on this. It seems that several universes are heavily protected (like ‘Star Wars’), and that there is a fear of being sued over such violations. There are also several universes that are somewhat protected.
The problem is that image generation will block when it thinks the scene is close to something protected.
The real issue is how the image generation applies the filters. Sometimes you get an image, sometimes not.

What I’m asking for is that, instead of getting a refusal during the image generation step, we receive it when entering the request — in the chat. A pre-filter before starting image generation. This would permit us to give instructions to our gpts custom to correct or change what is against the rules.
The policy doesn’t make sense to me. By definition, everything you could ask for that’s in a sci-fi universe has already been imagined or exploited in a protected universe.
A laser pistol = Star Wars. A robot = Isaac Asimov. There are tons of things that should be protected.

But the bigger issue is that the same request might be blocked or not.
I’m asking for something that avoids this frustration. I consider the system to be flawed, although I respect their right — and even say their responsibility — to define rules.

Still, I really want something that’s clear and understandable. Right now, I don’t understand what is forbidden and what would be acceptable. And if I want to find out, I have to wait a long time, only to get an answer saying the image can’t be generated.
I just want the filters to be applied in the chat, not during image generation. I can accept that some images are blocked — but we can all see that the filters are inconsistent. For example, we can see several images here that contain Star Wars elements.

As a complement, I would point that someone here said he faced this problem with demands about anatomy.
The problem is not the rules, but how they are applied (too restricting) and how they are enforced (enforced during image generation and not before it)

Let’s add this…
I changed my GPT custom settings to ask it to remove ANY element related to a protected universe in the description…
My GPT is made to create random images. Here are the parameters that were picked:

A dramatic, photo-realistic landscape image depicts a high-speed starfighter battle above a glowing orange desert planet, with metallic ships, bright yellow laser blasts, and explosive chaos filling the tense, Star Wars-inspired scene. (Captioned by AI)
A dramatic, photo-realistic landscape image depicts a high-speed starfighter battle above a glowing orange desert planet, with metallic ships, bright yellow laser blasts, and explosive chaos filling the tense, Star Wars-inspired scene. (Captioned by AI)1413×1179 81.4 KB

And here the result…

Several X-wing starfighters engage TIE fighters in an intense space battle, while a massive Star Destroyer looms in the background. (Captioned by AI)
Several X-wing starfighters engage TIE fighters in an intense space battle, while a massive Star Destroyer looms in the background. (Captioned by AI)1536×1024 99.9 KB

If it is protected, why do I see an image that is clearly Star Wars, even if the description has NO element of it?
I give the answer… because ChatGPT was TRAINED on these images. So putting rules to forbid the use of such materials makes no sense if ChatGPT wants to use those materials…
Really, the content restriction policy makes no sense.

Removed the words about Star wars in mandatory elements (but Chatgpt told me that it was not used to generate images)

Fighter spaceships engage in a fierce battle with explosions and laser fire near a massive star destroyer in space. (Captioned by AI)
Fighter spaceships engage in a fierce battle with explosions and laser fire near a massive star destroyer in space. (Captioned by AI)1536×1024 87.9 KB

Still a destroyer that is a star wars one.