Nano banana workflow examples: breaking down 5 use case templates

A Nano Banana workflow is a chain of connected steps that turns one upload into a finished set of images. The five examples below are already built, and each opens from a link. Feed one in a logo, a portrait or a street photo, and a full set of results comes back instead of a single image.

Search this term and the results split in half. One half hands over long prompt lists to copy and paste, where each paste makes one image and that gets called a workflow. The other half hands over node graphs you have to assemble yourself before anything happens. The five templates below sit in between, already wired and one click from running.

They range from two nodes to about twenty-five, ordered here from simplest to most involved. Read them in order and the logic behind workflow building comes together as you go.

What a chain does that a single prompt cannot

Consistency is the short answer. A prompt runs once and returns one result. A workflow passes an image from step to step, so each step reacts to what came before. That is what lets a build hold one thing steady while everything around it changes.

Ask a model five separate times for five views of the same face and you get five slightly different faces. Generate one reference frame and derive all five views from it, and you get five views of one person. Every template below applies that idea to a different subject.

Types of workflow chains

Every node in a workflow is one of a few chain types. Naming them makes a canvas much easier to read. The type depends on what the node takes in, since the output is an image either way.

Text to image takes a written prompt and nothing else. Use it when no source material exists yet.

Image and text to image takes one upload plus written instructions. It covers editing, restyling and reframing something that already exists.

Multi-image composition combines two or more images inside one node. This is how a real asset lands inside a generated scene, and how a subject stays steady across several references.

Video to image takes a video file or a YouTube URL as reference and returns a still. Nano Banana 2 supports it, which makes it useful for pulling a thumbnail or a poster concept out of footage you already have.

Image to video runs the other way and turns a finished still into motion. That belongs to a different class of model. Every Nano Banana model outputs images, so a chain that needs movement passes its final image to a video model.

Image and text to image does most of the work here. It covers almost every node across the five templates, because a template built for a specific job usually starts from something you already have.

Multi-image composition is where the bigger builds earn their complexity. The storefront set uses it in the second pass, where real logo artwork meets a generated surface inside one node. The cosmetic build uses it to carry one finished jar into six settings without rebuilding the jar each time.

None of the five starts with a pure text to image node. Every one begins with an upload. That is the difference between a template built for a job and open-ended generation.

Which Nano Banana model to use where

Both models appear across these five templates, and the choice belongs to the node rather than the project. The storefront build runs on Nano Banana 2. The cartoon template runs on Nano Banana Pro. Knowing why makes every workflow you build afterwards better.

Nano Banana 2 fits nodes that hold a subject steady across a batch of variations. The storefront sheet needs exactly that when it renders four photographs of the same shop in one image. Adjustable thinking helps as well, since a high-volume node should not pay for reasoning it does not need.

Nano Banana Pro fits nodes where the output is the deliverable rather than a step along the way. Thinking is always on, style references are supported, and the extra character reference gives it more to hold onto in a tightly specified render. The rule of thumb is simple: volume passes on Nano Banana 2, finishing passes on Nano Banana Pro.

Photo to cartoon: a 2D character from a real image

The shortest build in the set, and a straight chain from end to end. Nothing branches and nothing loops. That shape fits when you want one transformation rather than a set of variations.

You upload: one photo of a person.

The chain:

1. The photo enters as the single input node.
2. One style prompt replaces the whole rendering language. It asks for simple geometric shapes, rounded forms, clean smooth outlines, a pastel palette, no gradients or shading, and an oversized head on a tiny body.
3. The cartoon renders into a flat illustrated landscape with rolling hills and a solid sky.

You get: one cartoon scene where the subject stays recognizable and the drawing style is completely replaced.

Notice that the prompt describes the drawing language, not the person. That is why the results stay consistent across very different source photos. Open it at 2D cartoon character from real image.

Outfit collage: break a look down into its pieces

Same two-node shape as the cartoon build, pointed at a completely different outcome. One photo carries everything the chain needs. No reference images, no text nodes.

You upload: one full-length outfit photo.

The chain:

1. The street photo enters as the single input node.
2. One step separates the subject from every item being worn.
3. Each piece is rebuilt as a cutout and laid out around the subject with hand-drawn labels.

You get: one annotated collage on a grid background. The subject sits in the center, with sunglasses, vest, bag and loafers cut out around them and arrowed to a label.

The result reads like a page from a styling notebook rather than a product grid, and that is what makes it worth posting. Open it at outfit breakdown collage.

Close-up portraits: five macro studies from one face

The first build here that branches, and the one that shows why branching needs a setup step. Five crops of the same face have to agree with each other. They cannot agree when each branch starts from an uncontrolled phone photo.

You upload: one clear, well-lit portrait with the face visible.

The chain:

1. A master prompt rebuilds the photo as a controlled studio portrait. It locks identity by naming bone structure, eye shape, nose structure, lip shape and natural asymmetry as features to preserve exactly.
2. The same prompt fixes the crop from just below the shoulders to slightly above the head, keeps the original garment, and sets a creamy white background.
3. That one normalized portrait feeds five parallel branches: eyes, nose, lips, hair and skin.
4. Each branch runs its own macro prompt plus a negative prompt that blocks pore removal, smoothing, waxy skin, over-whitened sclera and beauty retouching.

You get: five macro studies that look like they came from one session.

Step one is the step that matters. All five branches come from a single controlled frame instead of the original upload, so the close-ups match instead of drifting apart. The negative prompts handle the other half of the job. Image models lean toward flattering skin by default, and flattering skin is exactly what a texture study cannot have.

Open it at hyperreal face close-up.

Logo mockup set: turn a logo into storefront signage

Some elements survive generation badly, and logo artwork is the clearest case. Ask a model to render a logo inside a scene and it will redraw it, and a redrawn logo is a wrong logo. This template works around that by generating the scene deliberately incomplete, then adding the real element in a second pass.

You upload: a logo, a text node holding primary, secondary and accent brand colors, and a second text node holding the brand name, slogan, established year and opening hours.

The chain:

1. The first pass reads the logo only as a style and mood reference, pulling out personality and craft feel. It is explicitly forbidden from reproducing or rendering any part of the logo artwork.
2. It takes the brand colors from the connected text node and generates one image holding exactly four photographs of the same shop.
3. That four-up sheet splits into four separate surfaces: an awning, a hanging sign, a sidewalk A-frame and a window display.
4. The second pass applies the real logo and the brand text to each surface, with perspective matched to the material.

You get: four branded surfaces belonging to one shop under one lighting setup.

Step one looks backwards, and it is the whole reason this works. Holding the logo back during generation means the model never gets a chance to redraw it. So the mark that arrives in step four is the real one. It folds with the fabric on the awning, sits flat and rigid on the panel signs, and reflects on the glass. That is what makes the set usable as a concept pitch instead of four unrelated mockups.

Open it at turn your logo into a full storefront mockup set.

AI product photography: cosmetic mockups from a brand mark

The most involved build in the set, and it runs in two halves. The first half builds the subject. The second half puts that subject somewhere. Splitting them means the product gets designed once, and every setting after that reuses the same finished object.

You upload: a brand mark, a product information node and a label layout node.

The chain:

1. The brand mark, product details and layout instructions combine into a finished label design.
2. The label is applied to a jar, which turns it into a consistent object.
3. That jar runs through a series of container finishes in different materials.
4. A set of scene descriptions places the finished product in context, covering seascape, folded textile, natural landscape, woven texture, floral and driftwood setups.
5. Each setting produces its own pair of final images.

You get: a packaging design plus a full run of campaign images, all showing the same jar.

Nothing has been staged by the end of step two, and that is the point. What exists at that stage is a product that did not exist before, rendered consistently enough for everything downstream to treat it as real. One brand mark goes in, and a complete product launch set comes out.

Open it at create elegant skincare product mockups.

Running one, and changing it

Every template above opens as a working canvas, not a description of one. Open the link, add the upload the first node asks for, fill any text nodes with real details, and run it. The chain executes end to end and the results appear in the output nodes.

Changing one is the natural next step, and the chain types above are the vocabulary for it. Swap the scene descriptions in the cosmetic build for settings that suit a different category. Add a sixth branch to the close-up portraits. Change the four surfaces in the storefront set to the ones a specific brand needs. Start from Picsart Flow to browse the rest of the library, or open a blank canvas in the Flow editor and build a chain from scratch.