Image to Prompt Tutorial: Reverse a Reference Style, Then Reshape It in NanoBanana Pro

You have a reference image you genuinely like, but you cannot say why it works. Feeding it straight into image-to-image and hoping for the best usually returns a batch of results that share the colour and lose the character. The cheaper path is to describe the image first, pulling composition, lighting, colour and texture out of the frame and into sentences you can edit, store and reuse. That step is what image to prompt means in practice.

This article covers three things. First, the three-step reverse workflow. Second, what each of four common scenarios actually needs from a reverse prompt. Third, how to hand the result to NanoBanana Pro for image-to-image. The case images below come from publicly documented reverse-prompting tool demos, so you can reproduce them.

What image to prompt is, and how it differs from feeding the reference directly

Feeding the reference directly means putting the reference and the target into image-to-image together and letting the model move the style across. That approach gives you one strength dial to turn. What the model absorbs is usually the surface colour distribution, while the composition and lighting logic stay behind. Image to prompt inverts the order: ask the model to explain in words how the picture is built, and the style becomes a set of parameters you can edit line by line.

Approach	What you control	Drift risk
Feed the reference directly	One style strength value	High, colour only
Reverse the prompt first	Lighting, lens, composition, palette	Low, style and subject stay apart

The difference is data structure. Once an image is inside the model, style and subject are fused. Once the style is written down, style is style and subject is subject: you keep the lighting and lens language and replace only the people and objects, and swapping the subject costs one line of text. That separation is the real value of reversing a prompt.

The three-step workflow, from one reference image to a reusable prompt

Step	What you do	What you get
1. Upload the reference	Drop in the photo, render or illustration you want to reproduce	The visual information the tool can read
2. Name the focus	Say whether lighting, palette, composition or lens matters most	Attention concentrated where you need it
3. Copy and refine	Take the prompt and push back with one more sentence	A piece of text you can reuse

The workflow itself is short. Step two is where results are won or lost. These three habits decide whether the first reverse prompt is usable as-is.

Name the focus instead of asking for an analysis

Ask for an analysis of the image and you get a description that touches everything and commits to nothing. State the priority and the output changes shape. If texture is the goal, ask for a breakdown of lighting and colour grading. If you are studying structure, ask for composition and lens. The same reference image produces a differently weighted prompt depending on what you asked for.

Say what the prompt is for, and the wording adapts

Tell the tool where the prompt is going. Saying it is for a product shot pushes the material, angle and lighting detail up. Saying it is for an anime scene shifts the description toward rendering style and atmosphere. Declaring the destination narrows the vocabulary the tool reaches for.

Push back with one sentence

A reverse prompt is a draft, not a verdict. Too dense, answer with a request to simplify. Too flat, ask for more cinematic treatment. The tool rewrites that section on the spot. Explaining the intent and letting it rewrite usually beats hand-deleting words yourself.

Real case: the full prompt reversed from one portrait

The reference portrait below is unremarkable in content. It is one person beside a window. Everything that makes it work lives in the lighting and the lens, which is exactly the kind of image image to prompt handles best.

Hand this image to a reverse-prompting tool and what comes back looks roughly like this.

Realistic cinematic editorial portrait. [SUBJECT: a young woman with light brown hair pulled back neatly, hazel eyes, balanced natural features, a calm and gentle expression with the faintest hint of a smile], wearing [WARDROBE: a simple off-white or cream top]. Tight head-and-shoulders framing, eye-level camera, face centred with only a slight turn into a three-quarter angle, looking straight into the lens. Natural understated makeup, muted nude lip, smooth translucent skin with realistic fine texture and a healthy natural sheen. Soft warm diffused window light entering from frame left, wrapping the face gently and leaving delicate highlights across the forehead, nose bridge and cheekbones, with low-contrast soft shadow on the right. Neutral indoor background, fully dissolved into a creamy beige bokeh. Very shallow depth of field with the eyes held sharp; full-frame camera, 85mm portrait lens, aperture around f/1.8 to f/2.2, refined cinematic grading, clean and elegant frame with editorial magazine finish.

The part worth studying is the bracketed pair, [SUBJECT] and [WARDROBE]. Those slots grew out of the prompt itself. Swap what is inside the brackets for a different person and a different garment, and the lighting, lens, depth of field and grading outside the brackets stay untouched. The same visual system transfers to a new subject.

That is the real gap between a reverse prompt and an ordinary caption. A caption lets you recognise the image. A prompt with slots lets you produce the next image in the same visual system.

Four scenarios that get reversed most often

Reverse prompting is not limited to wanting a near-copy. Laid out side by side, these common jobs each need something different from the output.

Scenario	Reverse for	Then what
Reproducing a look you love	Lens, lighting, grading	Generate more subjects inside the same aesthetic
Matching an existing product shot	Angle, material, lighting	Extend the line into new variants without rebuilding the set
Turning a mood board into words	Palette, material, lighting, layout	Send the recipe to the whole team instead of guessing at the feeling
Understanding why a photo works	Lighting, lens, composition, colour	Apply the same reading to your own prompts

Reproducing a look you love

A composition someone sent you, a render you scrolled past, anything that reads as “I cannot explain it but I like it” is a candidate. The lens, lighting and grading come back as explicit parameters rather than a feeling.

Matching an existing product shot

You already have a set of brand product images and want to keep producing new items without rebuilding the set. Feed one of the existing shots in and take back a prompt describing angle, material and lighting, so the new item sits inside the same visual standard.

Turning a mood board into words

The problem with a mood board is that it cannot be handed over in sentences. Reverse prompting converts one reference into a written list. The team generates from the same recipe and the result does not wander depending on who ran it.

Understanding why a photo works

Breaking one image into lighting, lens, composition and palette is a free lesson in visual craft. Once you can read the decisions behind it, you know where to start when writing your own prompt.

Five prompt dialects, and why the wrong one wastes the work

The same scene needs different wording for Midjourney than for Stable Diffusion. Third-party reverse-prompting tools document their outputs as separate dialects precisely so that Midjourney phrasing does not get pasted straight into a local workflow.

Prompt dialect	Feeds into	Output character
General	Any generator	One or two complete sentences, most portable and most conservative
Midjourney	Midjourney	Stacked short clauses with aspect and version direction attached
Stable Diffusion	Stable Diffusion, ComfyUI	Tag-style phrasing with optional weights, positive and negative split apart
Flux	Flux family	Full sentences, with subject, spatial relations and surface quality described in detail
Nano Banana	Nano Banana family	Natural-language description that states which visual information should be kept or changed

The choice follows where you will actually generate, not which dialect sounds most advanced. The Image to Prompt tool on this site makes the target platform a pre-generation option covering Nano Banana, GPT Image, Midjourney, Stable Diffusion and Flux, and the wording shifts accordingly. Output language can be set to English, Simplified Chinese, Traditional Chinese or Japanese.

Handing the reverse prompt to NanoBanana Pro for image-to-image

Reversing gives you the style. Image-to-image is what applies that style to your subject. For a stable result, organise the reverse output as a JSON style contract instead of pasting prose.

First, have Gemini extract structured visual information from the reference:

Extract the visual style from the input image and return strictly valid JSON only, with no surrounding text. Suggested fields: colors (palette, dominant, accents, tone_contrast), composition (layout, balance, focal_points, spacing_and_rhythm), visual_effects (textures, lighting, shadows, filters), overall_style (design_language, emotional_tone, reference_genres). Leave fields empty when the image does not support them rather than inventing values.

Second, treat that JSON as the style contract and hand it to NanoBanana Pro along with the image you want reshaped:

You are an image director. Reshape the uploaded image into the same aesthetic system described by the JSON below, keeping the subject structure and semantics unchanged unless I say otherwise. JSON: {…}

The payoff shows up during iteration. JSON splits style into individual fields, so testing a different light means editing lighting alone and changing the palette means editing colors.palette alone. One change at a time is the only way to know which edit caused the shift. When style information is buried in prose, changing one word drags the tone of the whole paragraph with it and comparison becomes impossible.

One practical warning before you start. Subject identity, in-image text, brand marks and cultural symbols are the elements most likely to be misread. Compare those against the source image before generating, and do not accept the reverse output wholesale.

Frequently asked questions

How does image to prompt actually work

The tool analyses the visual content of the uploaded image, covering composition, lighting, colour, style and subject, then returns a detailed written prompt that lets a generation model reproduce a similar effect. It is not identifying what happens to be in the frame; it is writing down why the frame works as an executable instruction.

Can I use the reversed prompt directly

Use it as a starting point rather than pasting it verbatim. A reverse prompt is a new prompt rebuilt from visible evidence, not a recovery of the exact wording the original creator used. Pick the sections you need, replace the subject and brand-specific language, then test it in the model you actually generate with.

What do I do when the reversed prompt is too long

Keep the two or three fields closest to your goal. Editing light means keeping visual_effects; changing structure means keeping composition; delete the rest. Fewer fields usually means better compliance.

The subject deforms when the style is too strong. What now

Add a hard constraint to the instruction, such as preserving the subject silhouette and identity features, or leaving facial structure unchanged. Subject and style are separate things, and the model needs the boundary stated explicitly.

Why does the colour keep drifting

Pin colors.palette and change one colour range at a time. Testing two ranges at once makes it impossible to tell which one caused the shift.

Which dialect should I pick

Decide which model will do the final generation first, then pick the matching phrasing. Midjourney phrasing for Midjourney, tag-style phrasing for Stable Diffusion or ComfyUI, Flux phrasing for Flux, and the general dialect everywhere else. Picking the dialect before the model usually means doing the work twice.

Are uploaded images stored or shared

An uploaded image is used as the input for that analysis task only, and content may be processed through third-party AI infrastructure with retention depending on functional and operational requirements. Do not upload private, confidential or rights-restricted material you have no right to use, and crop or mask screenshots containing personal data, account details or business secrets before uploading.

Can the reversed prompt be used on other platforms

Yes. Each dialect exists for a specific platform, so Midjourney phrasing pastes straight into Midjourney and tag-style phrasing suits ComfyUI or Automatic1111. When moving across platforms, recheck aspect ratio, negative prompting and reference settings, because different models read the same wording differently.

Start with the Image to Prompt tool on this site

What image to prompt really saves is trial and error. Instead of staring at an image guessing why it works, let the tool write the lighting, lens, composition and palette out as text, and limit your own job to deciding what stays and what changes.

The tool on this site accepts a single JPEG, PNG or WEBP upload. Before generating you can set the target platform, output language, detail level and visual focus, with an estimated point cost shown up front and the actual cost displayed once the task completes. The structured result lists subject, style, composition, lighting, camera, colour palette and aspect ratio as separate sections so you can take what you need piece by piece.