Image generators remember more than the output.

A generation request reads as a transaction — input in, image out, nothing left behind.

The prompt is the part that doesn't leave.

It stays attached to an account, accumulates across sessions, builds a record the user never explicitly authored.

The model learns the aesthetic before the user names it.

I hadn't thought about what that implied until a platform surfaced my style preferences back to me.

Prompts from three weeks earlier, repackaged as a profile. The granularity was specific enough that I recognized the pattern — but I hadn't constructed it consciously.

It assembled from requests I treated as separate, unrelated, temporary.

A behavioral signature — not an image.

OpenGradient Chat kept surfacing around exactly this question — how the infrastructure handles the prompt before the model sees it.

Image Studio in OpenGradient Chat sits behind TEE infrastructure — the operator cannot read the prompts.

Not something the operator chose. The architecture doesn't give them access to choose. Gemini, ByteDance, xAI — three different model integrations, same condition underneath.

The session ends. Nothing that connected the prompt to an account stays behind.

Whether that holds under real load across all three, I haven't verified independently.

The attestation covers the gateway.

What happens before the prompt arrives — whether the enclave assumption stays intact across updates.

Nobody has published a test for that condition yet.
$OPG #OPG @OpenGradient