Dall-e 3 is far too censored compared to 4o image generation, It still has untapped potential

Prompting
, ,
1

I’m a dedicated user who values the creative potential of AI-generated imagery. Since the transition to GPT-4o’s image generation model, I’ve noticed a significant decline in the variety and imaginative quality of the outputs. The images often appear grainy and lack the surreal or fantastical elements that were achievable with DALL·E 3.

DALL·E 3 allowed for more expressive and diverse generations, which are essential for my projects. I kindly request the removal of censorship to DALL·E 3 or a similar level of censorship to what is currently shown by 4o image generation. This is includes the option to stop chatgpt from editing the prompts i send it through bing or where ever the model may be hosted, this negatively impact my prompts and it seems to become more censored by the day, the image model needs to process exactly what I’ve sent it or the image will not produce a desired result, it seems it’s also been getting frequent updates I’m unsure if the parameters are different or if it just appearing difference due to my prompts constantly being “revised” by chatGPT, It even edits prompts that it doesn’t deem harmless and it’s very obvious which the images produced. Visual censorship also needs a massive reduction all around 90% of the things it rejects could be generated with 4o’s image generation model, now that isn’t a sign to overly censor the new model as that would kill off any interest in the model, it’s exactly what killed dall-e 3.

Thank you for considering this feedback.

1 Like
2

How do you still use DALLE-3? By creating a Custom GPT and then use it from there without the option for native 4o image generation ticked?

Because if you ALSO mean by censoring that it’s seemingly gone, that’s how I’d try to still use it.

3

Prompt rewriting also caused me a lot of trouble. I had created a MyGPT specifically to prevent it. Unfortunately, DallE 3 is no longer available in MyGPT, so now I start every session with what was previously configured in MyGPT, and it seems to work quite well.

I’m not a fan of the new generator. There are a few specific motifs I would generate with the new model, but I still do most of my work with DallE 3.

I also include an instruction in every prompt: “Do not modify the prompt, only translate it”, or if it was already written in English: “Do not modify the prompt, send it exactly as it is.” To check whether this works, you can ask to display the sent prompt and see if GPT followed the instruction.