Nano Banana Pro vs ChatGPT Image: Furniture Test
Our own Showcase test: Four furniture products, two models, 16 outputs. Jump to methods and images.
A beautiful living room is of little use if the sofa in it has changed. For product images, the first question is: Does the product still match your reference? Room styling, lighting and resolution come next.
Nano Banana Pro and ChatGPT Image offer different starting points. This guide helps you choose one and judge the result against your own product. Coming from the model picker? Jump to Settings in Showcase for formats and resolution.
Which model should I choose?
| Your main goal | Start with | Check carefully |
|---|---|---|
| Preserve the view of your product cutout | ChatGPT Image | Check product details and room perspective separately |
| Develop a styled room and creative scene variations | Test both | Check for unwanted rotation or changes in shape |
| Generate an image at specific pixel dimensions | ChatGPT Image | Showcase applies edge-length and aspect-ratio limits |
Our recommendation after the new series: When the reference product view matters, try ChatGPT Image first. Check every attempt: ChatGPT changes the sofa viewpoint too. For room styling, compare both yourself; we have not established an overall creativity ranking.
Why product fidelity comes before a beautiful scene
A small change can make a sectional sofa depict the wrong variant: the chaise moves to the other side, an armrest becomes rounder, or three seat cushions become two. A shelf can acquire extra compartments; a table can gain different legs or a thicker top.
Check view, silhouette and structural features first. Then inspect upholstery, wood grain, seams, colours and hardware. Finally, assess scale, contact with the floor and shadows.
A higher output resolution does not fix these changes. A sharp image of an altered product is still unsuitable for the listing.
Advantages, disadvantages and similarities
| At a glance | Nano Banana Pro | ChatGPT Image |
|---|---|---|
| Advantages | Creates uncluttered rooms in this series. Retains the cabinet's frontal overall view in both attempts. | Stays closer to the armchair and table reference views in both attempts. Also supports custom pixel dimensions. |
| Disadvantages | Changes the product viewpoint more often in our examples. No custom pixel dimensions in Showcase. | Its second sofa attempt also changes the view. Product details and room perspective still require inspection. |
| Similarities | Both create lifestyle images from product cutouts and offer 1K, 2K and 4K in Showcase. Both can change product details or obscure them with added decoration – check the result against your reference. | |
Nano Banana Pro: developing the scene
Nano Banana Pro belongs to Google’s Gemini family. Google positions it for complex visual tasks and precise creative control. Source: Google’s model overview
In Showcase, it is a useful starting point for exploring room styles, decoration and lighting. Evaluate the scene and the product depiction separately.
November test observation: Pro substantially changes the view in both sofa attempts, the second armchair attempt and the first table attempt. The cabinet remains frontal in both attempts. A convincing new camera angle can still violate the product-preservation brief.
Another model option: Supported modes also offer Nano Banana 2 (“Nano Banana · Fast”). The image examples here compare Nano Banana Pro and ChatGPT Image.
ChatGPT Image: product view and custom sizes
The new series uses the visible Showcase model name ChatGPT Image. Earlier documentation calls the integration Sunburst; the test dialog did not expose a specific endpoint version.
ChatGPT stays closer to the original view in both armchair and table attempts and the first sofa attempt. Its second sofa attempt is more frontal. This supports trying it first for an established product view while continuing to check every output.
This is not a promise of pixel-perfect preservation. Upholstery, edges, proportions and small parts still need inspection. An action such as rotating a product deliberately asks for a new view; retaining the reference angle would not be the right success criterion there.
Another benefit in Showcase is custom sizing. You can specify a width and height, such as 1600 × 1200 pixels, when your layout already has fixed dimensions.
Furniture test: four products, 16 results
On 3 November 2026, we tested a corduroy sofa, a tall wooden cabinet, an upholstered armchair and a wooden dining table in Showcase. Each product received two attempts with Nano Banana Pro and two with ChatGPT Image. All 16 outputs appear below, including the less suitable versions. The room images shown are AI-generated.
How we tested
| Test condition | Setting |
|---|---|
| References | Four existing product images; the same file per product for both models |
| Action | Milieu Creative, “Möbel-Marke” brand template |
| Scene | Natural Daylight / Contemporary Living Space |
| Format | 1:1, 1K; every delivered file actually measures 1024 × 1024 pixels |
| Additional options | No avatar or extra detail reference; True-Tone and product lights off; no product dimensions supplied |
| Repetitions | Two per model and product; no pilot images, retries, retouching or upscaling rounds |
| Order | Alternating models; sofa/armchair started with Pro, cabinet/table with ChatGPT |
The interface displayed “Nano Banana Pro” and “ChatGPT Image”, without a specific endpoint version. This test describes the Showcase integration on that date. The chosen brand template, mode and scene settings are part of the process; the entered prompt alone does not fully describe it.
Show the exact prompt – identical for both models, in German
Setze das Produkt aus dem Referenzbild in ein helles, modernes Wohnzimmer mit hellem Eichenboden, neutraler Wand und weichem Tageslicht von links. Erhalte Ansicht, Ausrichtung, Form, Farbe, Material und alle sichtbaren Produktmerkmale der Vorlage. Zeige das Produkt vollständig und unverdeckt. Ergänze nur die Umgebung und einen passenden Bodenschatten. Stelle keine Gegenstände auf das Produkt oder in seine Fächer. Keine Menschen, Schrift oder Logos hinzufügen. Fotorealistische Raumdarstellung mit einer zur Produktansicht passenden Perspektive.
We assess product identity, material appearance, spatial plausibility and adherence to the brief. Codex performed the initial visual review using the references and exports. There was no measured colour or dimensional accuracy and no blinded human assessment. Four products and two attempts per model support specific observations, but do not establish a general success rate. Runtime and credits were not used as comparison criteria.
Previews are resized for the web. The WebP downloads below the images retain the full resolution. On desktop, the models appear side by side for each attempt; on mobile, each Pro image is followed by ChatGPT.
1. Corduroy sofa: the second attempt matters too
The broad left outer side and visible seating surfaces reveal the reference camera angle. Pro makes the sofa more frontal in both attempts. ChatGPT initially stays closer to the reference, but its second attempt also turns more towards the front. ChatGPT does not guarantee preservation of the view either.
Pro makes the sofa substantially more frontal in both attempts. ChatGPT stays closer to the angled reference in attempt 1; its second attempt also becomes more frontal. Two back cushions and the right-hand chaise remain recognisable. Armrests, seat contours and feet need a detail check in every version.
Select an image for full resolution as compressed WebP.
2. Tall cabinet: both stay frontal
This case does not support a blanket claim that Pro rotates products: both models keep the frontal overall view in both attempts and leave the cabinet empty. Their rendering behind the glass door differs substantially. Preserving the view and preserving product details require separate checks.
Both models retain the frontal overall view in both attempts and add no decoration inside the cabinet. Pro makes the area behind the glass much darker; ChatGPT renders it lighter. Wood grain, frame details and proportions do not fully match the reference. A frontal view alone does not establish an identical product depiction.
Select an image for full resolution as compressed WebP.
3. Upholstered armchair: check buttons, armrests and base
Every version shows four backrest buttons and two armrests. Pro’s second attempt makes the chair almost frontal; both ChatGPT attempts stay closer to the angled reference. Counting the visible base parts is insufficient: their angles and connection to the seat also need to match.
Four buttons and two armrests remain visible in all four images. Pro switches to an almost frontal view in attempt 2, changing the visible arrangement of the base. Both ChatGPT attempts stay closer to the angled reference. This does not establish exact preservation of upholstery shape, base or fabric texture.
Select an image for full resolution as compressed WebP.
4. Wooden table: Pro does not change the view every time
Pro’s first attempt creates a noticeably angled view. Its second attempt is substantially closer to the reference. ChatGPT retains the wide overall view in both attempts. A single attempt would have made the difference between the models look larger here. Both change the individual wood grain.
Pro creates a strongly angled table view in attempt 1. Its second attempt is substantially closer to the reference. Both ChatGPT results retain the wide, approximately frontal overall view. In all four outputs, inspect the top grain, edge thickness and overlap of the black base parts individually.
Select an image for full resolution as compressed WebP.
What these examples teach us about furniture images
For an established product view, these examples would lead us to try ChatGPT Image first. Sofa attempt 2 still reveals a limitation. Nano Banana Pro keeps closer to the requested view in both cabinet attempts and its second table attempt. The specific input and repetition affect the result.
Both models require product checks. A matching silhouette establishes neither identical wood grain nor accurate upholstery and base details. Lighting, shadows and a convincing room do not replace those checks. Applications requiring exact material representation or reliable dimensions need additional controls.
Earlier examples from September 2026
The following examples supplied by the Showcase team remain available as supplementary evidence. They belong to an earlier series with unequal attempt counts and formats.
Open earlier comparisons of three sofas and a tall cabinet
The following examples were supplied by the Showcase team: three sofas and a tall cabinet. They include every supplied variant, with four Nano Banana Pro results and six ChatGPT Image results in total. Attempt counts and output formats differ in places; complete settings and repeated test series are unavailable for these examples. The observations concern these images and do not establish an overall success rate or model ranking.
These four examples show where the product is preserved and where it changes. The new viewpoint is particularly clear in the cabinet example. The grey sofa also reveals a Sunburst limitation: Preserving the product does not automatically create a spatially convincing scene.
Tall cabinet: does the front view stay frontal?
The reference looks straight at the glass door. Nano Banana Pro creates an angled view with the right side visible. ChatGPT keeps the frontal view more closely. Both add decoration inside the cabinet. If the brief calls for an empty cabinet, that also needs checking.
Original cutout
Nano Banana Pro
2048 × 2048 px
ChatGPT Image · Sunburst
2048 × 2048 px
Nano Banana Pro reveals the right side, creating a new view. ChatGPT stays closer to the frontal reference. Both add objects inside the cabinet, so its interior is also changed.
Select an image for full resolution as compressed WebP.
View prompt
Place the tall cabinet into a modern living room.
Green sofa: comparing perspectives
The armrests and visible seating surfaces make the perspective easy to compare: Nano changes the viewpoint more, while ChatGPT stays closer to the reference.
Original cutout
Nano Banana Pro
2048 × 2048 px
ChatGPT Image · Sunburst
1488 × 992 px
Nano uses a lower-looking viewpoint; ChatGPT is closer to the cutout view.
Select an image for full resolution as compressed WebP.
View prompt
Add a skandinavian room around the sofa
Grey sofa: better product preservation, difficult room perspective
The two patterned cushions provide a concrete comparison: ChatGPT keeps them closer to the reference, while Nano replaces them with light cushions and adds throws. Yet the ChatGPT result raises a separate question: does the elevated view of the sofa fit the view of the surrounding room?
Here, the combination of the high product viewpoint and the window and sideboard perspective does not feel entirely natural. Sunburst can preserve the product better while still struggling to build a room that fits its perspective. Check product features and spatial plausibility independently. A convincing room on its own may still be incompatible with the camera angle of the cutout.
Original cutout
Nano Banana Pro
2048 × 2048 px
ChatGPT Image · Sunburst
2048 × 2048 px
ChatGPT keeps the two patterned cushions and the visible product view closer to the reference. Nano replaces the patterned cushions and adds throws. At the same time, ChatGPT’s elevated view of the seats does not feel naturally aligned with the window and sideboard perspective. Product fidelity and spatial plausibility need separate checks.
Select an image for full resolution as compressed WebP.
Red sofa: three ChatGPT variations
Different living rooms, a similar product view: the ChatGPT variations keep the red sofa’s orientation closer to the cutout. Nano Banana Pro shows it more from the side. Look closely at the black feet, tufting and cushion arrangement – these details can change even when the perspective stays similar.
Original cutout
Nano Banana Pro
2048 × 2048 px
ChatGPT Image · Sunburst · 1
2048 × 2048 px
ChatGPT Image · Sunburst · 2
1024 × 1024 px
ChatGPT Image · Sunburst · 3
2496 × 896 px
The three ChatGPT results show different rooms and framing. Their overall view stays closer to the wide reference; Nano exposes more of the left outer side. Differences in individual feet and upholstery details still need checking. Output formats and pixel dimensions differ.
Select an image for full resolution as compressed WebP.
View prompt
Erstelle eine fotorealistische Milieuaufnahme des Referenzprodukts in einem hellen skandinavischen Wohnzimmer. Heller Eichenboden, offwhitefarbene Wände, natürliche Leinentextilien, ein dezenter Wollteppich und sparsam eingesetzte Keramik schaffen eine ruhige, wohnliche Atmosphäre. Klare Linien, großzügige Freiflächen und zurückhaltende Dekoration. Platziere das Produkt als zentralen Blickfang, vollständig sichtbar und mit realistischen Größenverhältnissen. True-Tone Licht Modus Hochwertige Interiorfotografie mit klaren Details, ausgewogener Belichtung und natürlicher Perspektive. Keine Menschen, keine Schrift, keine Logos hinzufügen.
Settings in Showcase
The following describes the new model-selection experience in Showcase. Provider APIs may expose different options.
Choose the model and start generating
Generate opens the settings dialog. Choose the model and available output options there. The Generate button inside the dialog starts the job. In workflows, choose models on the individual nodes.
1K, 2K or 4K?
Choose a size for its intended use: 1K for initial inspection, 2K for larger displays, and 4K when you need larger files. Check the requirements of your store or layout.
For Sunburst in Showcase, these tiers deliver a long edge of 1024, 2048 or 4096 pixels. 4K includes upscaling. Nano’s labels describe model-specific resolution tiers, not identical pixel dimensions across every aspect ratio.
Aspect ratios and custom dimensions
For regular generation tasks, choose from the available aspect ratios. ChatGPT Image also supports custom pixel dimensions: whole numbers from 64 to 4096 pixels per edge, with an aspect ratio between 1:3 and 3:1.
Find the right model for your catalogue
- Choose three representative products: one with clear geometry, one with a challenging material and one with small distinctive details.
- Use the same original cutout, references and task description. Keep the mode, aspect ratio and output tier consistent where possible.
- Generate the same number of results with each model. Keep every variant and record the actual settings. If you measure generation time, record when the test took place; queues can affect the duration.
- Mark unwanted product changes before judging the scene. A beautiful interior does not compensate for the wrong product variant.
- Compare which results you could use directly and which need correction. Check whether the same product details survive a second attempt.
Example test instruction: “Place the product from the reference in a bright living room with an oak floor and soft daylight from the left. Preserve the reference view, orientation, shape and all product features. Add only the surroundings and an appropriate contact shadow.” This is a suggested test brief, not the prompt of an existing result shown here.
Frequently asked questions
About the author
Author
Tim Hoffmann
Chief Product Officer, getshowcase.ai
Tim Hoffmann leads the product strategy for the AI image studio at showcase (getshowcase.ai). He brings years of e-commerce experience in product data, marketplace integrations, and visual content creation. His focus: helping Home & Living retailers turn product cutouts into photorealistic lifestyle images and room scenes in minutes - without expensive shoots, with measurably better conversion. Tim shares practical strategies for product images that perform on marketplaces and in your own shop.
