GPT Image 2 vs Nano Banana 2: Prompting for the Image You Actually Want

Those requirements are easy to lose once an impressive image appears. Consider a restaurant promotion: the fruit glistens, the light looks natural and the background has a lovely softness. But if that softness also obscures the slice you wanted people to inspect, the image has missed an important part of the brief. Adding more words about realism would not explain what needs fixing.

This is where prompting becomes a creative decision. You have to decide what deserves attention, which details establish the mood and which details must be exact. Then you can distinguish an appealing interpretation from a problem that needs another edit.

The examples below put those decisions into practice with a fragrance campaign, a stained-glass workshop, a restaurant photograph and a festival poster. Both tools make images worth considering. Their differences become much more useful when you look at what each image would help you communicate.

GPT Image 2 vs Nano Banana 2: What the Briefs Reveal

The OpenAI examples use ChatGPT Images 2.0; Google’s use Nano Banana 2 in Gemini. The images are labeled ChatGPT and Nano Banana 2 throughout so you can follow each result and its edits.

Each pair starts with the same prompt. Where a specific defect needed correcting, the original and one focused repair appear together. The final comparison gives both tools the same reference image, which makes it possible to judge how each handles a change to an already chosen composition.

Here are the choices the displayed results support. The discussion beneath each prompt explains why a different priority could change the decision.

Image You Need	Stronger Fit Here	Deciding Detail
Warm fragrance campaign with room for a headline	ChatGPT	Warm amber light and a dark upper-left copy area
Editorial opener that establishes a workshop	Nano Banana 2	Complete glass panel and more of the working space
Restaurant image that makes the filling appetizing	ChatGPT	Clearer cut surfaces and a successful fork correction
Expressive, asymmetric festival artwork	ChatGPT	Bold type balanced against the blue ribbon and smaller details
Festival poster with prominent practical information	Nano Banana 2	Larger, more conspicuous date and city
New cap material in an existing wide composition	Nano Banana 2	Closer preservation of the reference framing

A Fragrance Campaign Needs More Than an Attractive Bottle

Suppose the image is going behind a campaign headline. The product has to attract attention without occupying the entire canvas, and the headline needs a calm area where it can remain legible. That means the bottle’s position and the background are part of the design brief before any words are added.

The fragrance prompt makes those choices together. A bottle slightly right of center leaves room at the upper left. Dark water and distant cliffs provide atmosphere without adding another subject. Amber glass, bronze metal and wet black stone establish a restrained material palette, while the light is meant to make the glass and liquid visible.

This is a practical way to use subject, setting, action and composition in a prompt. Each detail should explain something the image needs to communicate. The words below are doing different jobs; they are not interchangeable ways of asking for quality.

Create a luxury fragrance campaign photograph, landscape 3:2. A single unbranded perfume bottle stands slightly right of centre on a low slab of wet black volcanic stone at the edge of a still tidal pool. The bottle is thick smoked-amber glass, a precise rectangular shape with softly rounded corners, filled halfway with amber liquid, with a brushed bronze cylindrical cap. Art direction: a narrow shaft of late-afternoon sun crosses the bottle diagonally, revealing the glass thickness and a believable refracted amber caustic on the wet stone. Behind it, a dark sea and hazy sandstone cliffs dissolve into soft focus. Restrained charcoal, bronze and warm honey palette. Low camera angle, 85mm product-photography feel, tactile glass and metal, tiny water droplets, carefully controlled reflections; bottle tack-sharp. Reserve the upper-left third as quiet dark atmospheric space for later campaign copy. Premium editorial advertising, confident and spare. Physically plausible contact, reflections and liquid level. No text, logo, extra bottles, floating objects or decorative flowers.

Look first at the upper-left corner, before looking closely at the bottle. ChatGPT keeps that part of the sky dark and comparatively quiet. The strongest warmth sits around the product and toward the right, so a pale headline could occupy the left without competing with the brightest light.

Nano Banana gives the scene a more subdued, overcast character. Its water and cliffs make a convincing coastal setting, but the gray sky is less strongly separated from the rest of the picture. This could suit a quieter fragrance identity. For the warm bronze-and-honey campaign requested here, ChatGPT’s treatment is the more persuasive starting point.

The material choices also hold together in ChatGPT’s image. Bright edges reveal the thickness of the bottle, the liquid glows, and reflections on the stone give the object a sense of weight. These are useful things to ask for when glass is the product: a dark transparent object can otherwise disappear into a dark scene, however attractive the background looks.

There is a more concrete problem in Nano Banana’s image. The brief asks for a cylindrical cap, but the closure has a visible spray opening. Changing the entire photograph would throw away a scene that already works, so the correction names that one defect and protects the composition around it.

Change only the bottle closure: replace the exposed spray nozzle with a completely closed, solid brushed-bronze cylindrical cap, with no spray hole. Preserve the cap location and approximate dimensions, the bottle geometry, amber liquid level, glass highlights, wet basalt, tidal pool, cliffs, lighting, caustic, camera angle, crop and upper-left quiet space. Keep everything else unchanged.

The corrected cap fits the cooler photograph. There is no need to ask for an entirely different coast just because one product detail was wrong. Equally, closing the cap does not make this the warm treatment originally requested; the lighting choice remains a separate decision.

Neither initial image delivers the literal narrow diagonal shaft of sunlight in the prompt. ChatGPT provides broad warm backlighting, while Nano Banana looks closer to soft daylight. If that beam were essential to a campaign concept, neither result would be finished. If the real requirement were warm, readable amber glass, ChatGPT’s interpretation could be acceptable.

Decide which of those requirements you mean before requesting the next change. Otherwise, you can spend revisions correcting a detail that does not matter to the layout while overlooking one that does. For an existing product, the actual bottle geometry and label would belong among the fixed requirements: supply approved product references and compare them closely, rather than relying on an invented bottle like this one.

The Workshop Shows Why Action and Framing Belong Together

A cinematic workshop can contain beautiful light and still say very little about the people inside it. The next brief gives the scene a specific moment: a restorer seats a piece of amber glass while an apprentice holds a lamp over the work. Their attention has a shared destination.

That relationship gives the lighting a purpose, too. The warm pool comes from a tool someone is using, and the cool light comes from the windows before dawn. The prompt connects the atmosphere to the activity, instead of leaving the generator to decorate a portrait with a dramatic glow.

Create a cinematic editorial photograph for a feature about craft surviving a changing city. Inside a small stained-glass restoration workshop just before dawn, an adult woman in her forties, wearing a faded indigo work shirt, carefully seats one amber glass piece into a leaded panel lying flat on a timber workbench. One adult apprentice stands opposite her, holding a small inspection lamp above the panel; both people are absorbed in the same task. Frame a medium-wide view from the bench corner at eye level: their hands and the glass are the focal point, their faces remain visible, and rain-streaked windows reveal a softly blurred city outside. A warm pool from the inspection lamp meets cool blue dawn light; show believable reflections, worn wood and fine glass texture, with restrained film grain and natural skin. Keep the scene quiet and observational, with exactly two people and physically plausible hand positions. Landscape 3:2. No lettering, logos, watermarks or collage.

Nano Banana’s wider view makes the workshop easy to understand. You can see the panel as a whole, the tools around it and the two people positioned on opposite sides of the work. That is valuable for an editorial opener about a craft: the setting helps explain what the article is about before a reader studies the details.

ChatGPT gives up some of that context for intimacy. The restorer’s hands and amber piece are much more prominent, and the reflections make the panel feel tangible. Both faces remain visible, although the apprentice’s hair extends beyond the top edge. The closer framing makes the act of placing the glass more absorbing, but it does not fulfill the medium-wide instruction as well as Nano Banana’s version.

These are two defensible pictures with different editorial uses. A feature about the workshop as a place benefits from Nano Banana’s context. A section about the care involved in restoration has more to work with in ChatGPT’s close view. Rejecting the second picture merely because it is tighter would overlook a useful option; accepting it as the requested wide opener would overlook the brief.

Make the Framing Requirement Concrete

The original prompt asks for hands and glass as the focal point, visible faces and a medium-wide view. There is room to interpret how much of the panel must remain inside that frame. If seeing the whole object matters, state that directly: keep the complete panel visible, with space around its edges and both people comfortably inside the picture.

That direction would clarify a future attempt. Photographic shorthand is only a starting point. A term such as medium-wide suggests a distance, while naming the things that must fit tells the generator what that distance needs to accomplish.

For this scene, I prefer Nano Banana as the opener. ChatGPT’s closer picture is compelling enough to keep for a different placement, rather than trying to force every successful image into the same use.

In Food Photography, Focus Can Change What You Are Selling

A restaurant photograph has to make the food appealing, but it also chooses what the viewer learns about the dish. An overhead view could show the arrangement of orange slices beautifully while revealing little of the cream underneath. The next brief therefore asks for a three-quarter view and two sharp cut surfaces: one in the tart and one in the removed slice.

The separate slice has a purpose beyond making the table look styled. It brings the pastry and filling closer to the viewer. Softening it would sacrifice some of the information that justified including it in the first place.

Create a premium restaurant editorial photograph of a freshly cut blood-orange and pistachio tart on a low ivory ceramic pedestal. The tart is the hero: crisp golden pastry, glossy overlapping blood-orange slices, pale pistachio cream visible in one neatly removed wedge, and a few natural crumbs. Place the removed slice on a small matching plate in the lower-right foreground, with its point facing the tart. Use a close three-quarter view at table height so the filling and pastry layers are visible; keep both cut faces sharp and let the background fall softly out of focus. Early-morning side light enters from the left, casting a long soft-edged shadow across dark oxblood linen and catching the orange flesh without blowing out the highlights. Include only one small silver dessert fork beside the plate; let the materials and appetizing texture carry the image. Portrait 4:5. No hands, flowers, scattered whole fruit, text, logos, watermarks or collage.

ChatGPT’s image puts the filling and pastry forward. The blood-orange flesh is a deep red, the cut faces remain readable and the darker background keeps attention on the food. It is the stronger choice when the promotion needs people to notice what is inside the tart.

Nano Banana suggests a more casual moment beside a cafe window. The softer light and visible room could be an advantage for an image about the experience of eating there. But the foreground slice is softer than the main tart, and its point faces away from the tart instead of toward it. Those departures weaken the specific arrangement the prompt was trying to create.

Notice that more saturated color alone does not settle this comparison. ChatGPT is doing a better job of presenting the requested cut surfaces. If the intended mood were a gentle cafe morning and the slice detail were secondary, Nano Banana’s atmosphere would deserve more weight.

A Small Prop Edit Can Leave a Bigger Problem

Both first images put the dessert fork on the plate, despite the instruction to place it beside the plate. That sounds like a modest correction. It is also a useful test of whether an edit can relocate an object while keeping the rest of the styling intact.

Both received this same follow-up:

Change only the fork placement: move the single silver dessert fork off the small plate and onto the oxblood linen immediately beside the plate. Preserve the tart, slice, plate, pedestal, cut faces, lighting, camera position, crop and all food textures. Do not add a second fork.

ChatGPT solves the placement problem. Nano Banana draws a fork in the requested destination but leaves another on the plate. The instruction explicitly rules out a second fork, yet the corrected image contains two.

ChatGPT’s correction is the usable choice from this pair, though small food and crumb details also change. That matters if you are choosing between versions of an invented restaurant concept; it matters much more if a client has already approved the precise presentation of a real dish. In the latter case, compare the food itself again, not just the prop you asked to move.

The Nano Banana failure also puts a limit on what a more careful prompt can promise. This was already a focused instruction with an explicit object count. Repeating it in increasingly emphatic language would not establish that the next result is correct. For a selected photograph with one misplaced utensil, a localized edit in a conventional image editor may be a better next move than regenerating food that already looks right.

A Poster Can Spell Everything Correctly and Still Need Art Direction

The festival brief asks for four short lines of English text, a strong typographic hierarchy and a cobalt ribbon. The fictional event is a design and sound festival in Kochi, so the artwork needs both a distinctive identity and practical information someone can find quickly.

Those are related but different jobs. The festival name can dominate the page; the date cannot become so inconspicuous that the poster stops working as an announcement. The ribbon adds movement, but it must leave the lettering readable.

Design a finished portrait poster for a fictional contemporary design and sound festival in Kochi. Make typography the main visual: large condensed black letterforms on warm uncoated ivory paper, with an electric-cobalt folded ribbon weaving through the open spaces between the letters. Use a rigorous asymmetric grid, generous outer margins and subtle screen-print texture; the ribbon should add depth while leaving every letter readable. The exact text, and the only text, must be these four lines in this order: "OFF GRID" / "DESIGN AFTER DARK" / "18-20 OCTOBER" / "KOCHI". Set "OFF GRID" very large across the upper half, "DESIGN AFTER DARK" as a smaller strong subheading beneath it, and the date and city as two crisp supporting lines near the bottom. Give each line a clear place in the hierarchy; do not add sponsors, tickets, websites or decorative microtext. Show only the flat finished artwork, not a poster hanging in a room. Portrait 4:5. No watermark.

Both posters reproduce the four requested lines of words and numbers in the right order. That gives us something more interesting to compare than a spelling failure. The hierarchy changes the character of the event.

ChatGPT’s enormous condensed letters and rougher print texture feel more forceful. The ribbon cuts across the upper area, folds into the page and reappears beside the small information block. That asymmetry gives the composition more tension, which suits the requested design-and-sound identity.

The cost is breathing room. The type crowds the outer edges, despite the request for generous margins. Its small date and city also demand more attention from the viewer. I prefer this as expressive festival artwork, but the margins and information size would need reconsideration for a placement where people have only a moment to read it.

Nano Banana’s layout is more centered and predictable. Its ribbon forms a large loop, and the date and city occupy much more of the lower portion. That loses some of the asymmetric character, but it is easier to find the practical details. For a small social placement where the date matters as much as the name, that is a meaningful advantage.

Specify the Relationship Between the Lines

The prompt does more than provide wording. It makes OFF GRID dominant, DESIGN AFTER DARK secondary, and the date and city supporting information. Separating those roles gives the model a hierarchy to work with. Exact quoted text and placement instructions are useful starting points, but supporting text can still end up much smaller than you intended.

If the date must remain readable in a narrow placement, describe it as a prominent information block rather than relying only on the word supporting. If the letters must respect a fixed safe area, give that space a clear role. These are decisions about where the poster will be seen, not requests for more visual sophistication.

The two results show successful short English copy; they do not settle performance on long passages or other languages. Even here, a final check should cover punctuation and spacing as well as words. For approved brand type, fixed legal copy or text that will change across versions, keeping the typography editable in a design tool is the more controllable route. A strong generated composition can still be valuable as the visual direction.

The Cap Edit Changes the Choice Again

After choosing a campaign image, you might want to explore one material without reopening the entire composition. A bronze cap and an ivory ceramic cap suggest different products, even when the bottle and scene stay the same. That is a reasonable refinement to ask of an image editor.

Both tools received the same original ChatGPT fragrance image below. Using a shared reference matters here: the question is how each changes this particular photograph, including the parts already worth keeping.

The prompt names the cap as the only change, then specifies the bottle, light, surroundings and framing that should remain. This follows the useful principle of stating both the change and what must stay fixed.

Edit the attached fragrance image. Change only the bottle cap from brushed bronze metal to matte ivory ceramic. Preserve the cap's exact shape, size and position. Keep the smoked-amber bottle, liquid level, all glass highlights, the bottle's position, wet basalt, tidal water, sandstone cliffs, sunlight, caustic pattern, shadows, camera angle, crop and upper-left copy space unchanged. Preserve the original image dimensions and composition. Do not add text, logos, objects or a new scene.

Nano Banana’s cap reads as smooth ceramic, and the photograph remains close to the original wide arrangement. ChatGPT also makes the material ivory, with more visible horizontal texture. But it returns a taller frame with more foreground, despite the instruction to preserve the composition.

For an isolated product experiment, the extra foreground might be acceptable. For a campaign layout built around the original wide image, it creates another job. The bottle, headline area and surrounding space now occupy different proportions, so replacing the original image could change the balance of the design.

Nano Banana is the better editor for this particular material variation. That result sits alongside its failed fork relocation, rather than cancelling it out. One tool preserved this frame more closely; the other successfully moved a prop in the food image. They are different editing problems, and the outcomes justify different choices.

This is also a reason to keep the approved original nearby while reviewing an edit. It is easy to concentrate on whether the new cap looks ceramic and forget the composition that made the image worth editing. The reference gives you a concrete basis for deciding whether the change is an improvement you can actually use.

Decide What Deserves Another Revision

ChatGPT is the starting choice from these results for richly lit product and food images where the hero subject needs to be prominent. Nano Banana earns its place when the surrounding context matters, as in the workshop opener, and it handles the shared cap-material edit more faithfully. Neither choice removes the need to look beyond the most impressive part of the picture.

There is also a point where continuing in the generator stops serving the asset. When the image is already chosen and the remaining work is exact typography, product identity or one small prop, use the process that gives you control over that remaining detail. The purpose of a better prompt is to bring the image closer to the brief; you still get to decide when an appealing interpretation is useful, when it needs repair and when another tool should finish the work.

For a repeatable comparison across more of your own tasks, use this process to test reusable prompts before switching models.