I’m reporting an ongoing issue with using reference images in ChatGPT image generation. Even when I explicitly instruct ChatGPT to pass the original reference images to the image-generation model, the results do not consistently follow those references.
I cannot verify from the chat interface whether the original images were actually passed to the image-generation model. Missing reference inputs, selection of a different image, and insufficient reference fidelity could produce similar symptoms, so I’m not claiming a confirmed technical root cause.
Expected behavior: use the explicitly selected reference images as visual inputs. If a requested reference cannot be accessed or used, clearly say so rather than silently substituting a text-only generation or claiming successful reference use without verification.
Could the team investigate this workflow and provide a reliable way to select and confirm which reference images were used for each generation? Guidance on the diagnostic information or reproducible test needed to distinguish an input-handoff problem from a fidelity issue would also help.
Related discussion (now closed): Chatgpt is not using uploaded picture as reference when generating new pictures
My description of the issue: ChatGPT - 參考圖傳遞問題
The shared conversation describes the concern; it does not contain generation logs or a complete failing reproduction.