Artists have always worked with references, but they do not all enter a work in the same way. A photograph may establish a face. A scrap of fabric may settle a colour relationship. A sketch may hold the composition while leaving the surface unresolved. Trouble begins when an image model is given all three and asked, vaguely, to �combine� them.

Nano Banana can read several visual inputs and continue an edit through conversation. That flexibility is often described as control. In practice, control comes from deciding which reference is allowed to speak about which part of the image.

A reference is not a mood board by default

Mood boards work because a human viewer can ignore what is irrelevant. A designer may look at a hotel photograph for its green and ochre palette without borrowing the furniture, architecture or people. A model needs that distinction written down.

Name the sources. Reference A holds the sitter's face and hair. Reference B supplies the red wool coat and brass buttons. Reference C is colour only. The prompt can then ask for a waist-up portrait of the person in A, wearing the coat from B, using only the palette from C.

The word �only� protects the border between influence and instruction. Without it, the palette image may also begin to reshape the room or the figure.

This is less glamorous than talking about inspiration, but it is closer to the actual work of making images. Selection, exclusion and hierarchy are creative acts. The model does not remove them; it makes their absence visible.

Editing asks a different question from generation

When an existing image already works, describe the difference rather than the whole picture. �Replace the vase with a clear bowl containing three green pears� is a usable change. Follow it with the details that have already been settled: the table, window light, crop and painted surface should remain as they are.

Restating the entire scene in new language gives the model permission to reinterpret it. That may be welcome during exploration. It is a poor bargain when the drawing, pose or spatial tension has already been found.

For repeatable studio work, a Nano Banana API can keep the prompt and ordered references attached to the same job. The order matters. A note that says �use Reference A for identity� becomes meaningless if nobody can later tell which file occupied that position.

The archive should keep rejected images too, at least through the life of the project. They reveal where a face began to drift, when a colour decision flattened the space, or which instruction accidentally made the image more literal.

Consistency is an artistic choice, not merely a technical one

A single image can absorb an accident. A sequence makes every accident answer to its neighbours.

Before generating a series, decide what repetition is meant to do. The same character might need an unchanged hairline and coat, while lighting and framing are allowed to evolve. A group of exhibition graphics may share type placement and a narrow palette but deliberately resist a single illustrative style.

This short �visual bible� should contain only the features the series needs to recognise itself. Too many fixed details can make every image feel like a variation of the same template. Too few can make the sequence look as though several unrelated commissions were placed side by side.

The model's ability to preserve a subject does not answer the more interesting question of whether the subject should remain stable. That judgment stays with the artist.

Style language deserves care

Asking for the signature manner of a living artist is both a weak brief and a way of avoiding description. It is more revealing to name the qualities at stake: dry brush texture, shallow space, compressed perspective, thin graphite edges, a restricted mineral palette.

Those observations may still lead to an image that feels familiar. They also force the maker to look closely enough to know what is being borrowed. The resulting prompt becomes an account of visual decisions rather than a request for borrowed authority.

Text inside the image needs similar precision. Supply the exact words in quotation marks and say where they belong. Then read the result as text, not as texture. A poster can be compositionally strong and still fail because one letter is wrong.

Motion begins where the still stops

An approved image can become the first frame of another work, but animation introduces time, action and a moving point of view. Those are new decisions. Passing the still into an image-to-video tool should begin with a motion brief, not a copy of the image prompt.

That separation also protects the image. If the face changes during a turn, the fault belongs to the moving sequence rather than to the original reference. If the still was never stable, motion will make the problem harder to ignore.

Nano Banana does not make reference work effortless. It makes the contracts between references unusually explicit. One image owns the face, another the coat, a third the palette. The art remains in deciding why those things should meet, and what must be left behind when they do.