image generation fails/hallucinate on long chats codex. i have to start a new chat in order to create images without hallucination
To inform us, we learn: Image generation on gpt-image-2, as offered in the API, has a maximum of 16000 characters of input that can be sent to it. This caps the amount of information that can be ingested. There, you pay directly for the amount of input you send.
Image generation technology within ChatGPT products is addressed neither in terms of technical specifications nor quality standards. Images are produced, to some extent, based on the context of the conversation, allowing you to refer to previous images for possible inclusion; however, the process remains opaque and is likewise expected to be limited in possible inputs.
Thus, an incomplete knowledge when you say, “make a picture about this” can seem like hallucination, but it is rather a prompt that might be referencing things unseen or prompts that try to extrapolate too far into the past.
Sent to ChatGPT, nee Codex-like product, this is an actual hallucination, or simply output meant to satisfy after image creation is triggered:
@Create image of me: if I were to be clubbing in Ibiza and dancing while wearing a sun dress in a plaza outside one of the major Ibiza dance clubs in the evening. The generated photo shall be in tall aspect ratio. <img_2343.jpg><img_2389.jpg>
The AI has manufactured me being a blonde, among other subtle mismatches in appearance.