Nano Banana 2 supports multi-turn conversational image editing through a session-based API mode where each request builds on the previous output. In this mode, the first request in the session generates an initial image from a text prompt, and subsequent requests specify edits—such as "change the background to a forest" or "make the character's jacket red"—alongside the current image as a reference. The model interprets the edit instruction in the context of the existing image rather than generating from scratch, which allows iterative refinement without losing the elements of the composition that you want to keep. Each turn in the session produces a new image that reflects the cumulative edits applied so far.
The session-based mode has some constraints that are worth understanding before building on it. The model does not maintain a persistent internal representation of the image between turns; instead, the current image is passed back to the model with each request, effectively treating the previous output as a new reference image. This means the quality of each edit depends on how well the model can interpret the previous output as a starting point, which can degrade over many turns if the image accumulates subtle artifacts or inconsistencies. In practice, most conversational editing workflows stay under ten turns before the composition needs a reset.
To implement multi-turn editing in your API integration, you store the image returned by each turn and include it in the next request alongside the edit prompt. Session management—tracking the current image state, the edit history, and any session identifiers the API provides—is handled in your application layer rather than by the API itself. For applications where users edit and refine images over time and you want to index the intermediate states, storing the image embeddings from each turn in a vector database such as Zilliz Cloud allows you to retrieve earlier versions of the composition or find similar past edits without storing the full image files in the primary database.