Generative AI Digital Signage: GPT Image 2.5 and Samsung AI Studio

What actually changed for digital signage content in 2026

Three things moved at once, and only one was a model release. Image generation got faithful enough to reference photos to become a production step rather than a toy. Generation moved inside the signage platform, so the person scheduling the playlist is now also the person making the asset. And video generation became cheap and portrait-native, which matters because most retail screens are vertical. One vendor exit is worth remembering too: OpenAI discontinued the Sora app on 26 April 2026 and the Sora API on 24 September 2026. Build the pipeline so a model can be swapped without taking your content library with it.

• •GPT Image 2.5 shipped on 8 September 2026 to ChatGPT, ChatGPT Work and Codex, with GPT-Image-2.5 Flare and Sunburst in the API.
• •OpenAI says Flare beats GPT-Image-2 on quality at 50% lower latency, and that Images 2.5 preserves reference-photo subjects more reliably.
• •Samsung VXT AI Studio turns static images into signage-ready video for Samsung's commercial displays, and won a Best of Show award at ISE 2026. Samsung says it is available globally in 1H 2026, with possible extra usage costs.
• •Google's Nano Banana 2, called through the Gemini API as Gemini 3.1 Flash Image, arrived on 26 February 2026 with better text rendering and in-image localisation.
• •Veo 3.1 Lite, announced 31 March 2026, makes 4, 6 or 8 second clips in 16:9 or 9:16 at 720p or 1080p.
• •Menu-board prices and stock counts are a data problem, not a generation problem.

The generative signage pipeline: brief to screen in six steps

A model is one step, not the pipeline. The teams getting value are not the ones with the best prompts. They are the ones who put generation between a locked brand template and a named approver, and who can say six months later which asset played where and who signed it off.

Which AI tool should you use for which signage job?

There is no single winner, because these four answer different questions. GPT Image 2.5 is a still-image model and the strongest choice when a real product photograph has to survive into a new setting. Samsung VXT AI Studio is the step after it, turning an approved still into motion inside the CMS. Google's stack wins when you need text inside the image in a regional language, or a portrait loop at the lowest cost per second. A designer with a stock library still wins on judgement. Prices below are vendor list prices on 24 September 2026; the qualitative columns are our assessment, not a benchmark.

What does an AI-generated signage asset actually cost?

Less than the argument about it is worth. Google publishes per-image and per-second prices, so the sums are simple. OpenAI publishes a token rate rather than a per-image price for GPT Image 2.5 and says plainly that token consumption differs by model and quality setting, so read the usage field on your own responses for a week before budgeting. As an upper bound only, OpenAI's table for its earlier GPT Image models puts a high-quality 1024x1536 portrait at 6,240 output tokens, which at the 2.5 rate of $30 per million would be about $0.19, roughly ₹18. All rupee figures here use ₹96 to the US dollar, the rate on 24 September 2026.

• •Gemini 3.1 Flash Image: $0.067 per 1K image, about ₹6, rising to $0.151 for 4K.
• •Gemini 3.1 Flash Lite Image: $0.0336 per 1K image, about ₹3, for high-volume drafting.
• •Veo 3.1 Lite: $0.08 per second at 1080p, so an eight-second loop is $0.64, about ₹61. Veo 3.1 Fast is $0.12 and full Veo 3.1 is $0.40 per second.
• •Worked example, illustrative: one outlet needing four stills and one eight-second 1080p loop a week, at five generations per approved asset, costs 20 x $0.067 plus 5 x $0.64, or $4.54, about ₹436 before anyone's time.
• •That multiplies by store only if each store needs different creative. Most networks need three or four regional variants, not forty.

Screen formats: why 1080x1920 is not a prompt

This is where most first attempts fall apart. A standard portrait panel is 1080x1920, but you cannot ask GPT Image 2.5 for that, because custom dimensions must be multiples of 16 and 1080 is not one. Generate at 1088x1920 instead, which satisfies the rule and stays inside the permitted aspect and pixel limits, then crop eight pixels. Veo 3.1 generates 9:16 at 1080p, which maps onto the same panel without cropping.

• •Portrait panel: generate 1088x1920 and crop to 1080x1920 rather than fighting the model.
• •Landscape: 1536x1024 and 1920x1080 both work; OpenAI flags resolutions above 2560x1440 as experimental.
• •Header strips and totems: Nano Banana 2 added 4:1, 1:4, 8:1 and 1:8 ratios in 2026, the first time these shapes were native.
• •Keep every price, legal line and logo inside the brand template, never inside the generated area. Our rule of thumb, not a standard: nothing load-bearing in the outer tenth of the frame.
• •Build a still fallback for every video asset, and remember Veo caps a clip at eight seconds, so a longer slot is a stitched sequence.

Rights, labelling and brand safety under India's 2026 rules

In-store screens are advertising, and India tightened up on synthetic advertising this year. On 12 May 2026 the Advertising Standards Council of India released draft guidelines for responsible labelling of synthetically generated content in advertising, aligned with the IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026 amended on 10 February 2026, open for consultation until 13 June 2026. The approach is risk-based rather than a blanket ban, but it draws a hard line exactly where signage operators are most tempted to cross it: product visuals.

• •High risk, prohibited even with a label: fabricated endorsements, exaggerating product results or features through visual representation, fake locations that look real, and using a person's likeness without consent.
• •Medium risk, labelling mandatory: synthetic influencers, replicating a real person's likeness or voice even with consent, realistic settings created entirely with AI, and demonstrating a product that does not currently exist.
• •Also medium risk: synthetic visuals of product performance, unless they replicate how the product actually performs.
• •Low risk, no label: decorative AI backgrounds, routine colour correction, obvious fantastical elements, and AI-drafted ad copy. ASCI suggests plain wording such as created using AI, but mandates no single label.
• •Our house rule: no generated faces of real staff or customers, ever, and no generated plate of food the kitchen cannot put on a counter.
• •OpenAI applies C2PA metadata and invisible watermarking to its images. Keep that metadata intact through your pipeline instead of stripping it on export.

An illustrative Diwali 2026 content calendar for screens

Diwali falls on Sunday 8 November 2026, with the five-day run from Dhanteras on 6 November to Bhai Dooj on 10 November. Generated creative changes the plan's shape: the slow part is no longer production, it is approval and translation. The calendar below is illustrative, built backwards from those dates for a retailer with a few stores and no in-house studio.

Mistakes that make AI signage look cheap

The failure modes are boringly consistent, and all are process failures rather than model failures. We see the same four on almost every network that adds generation without adding an approval gate.

How Tech Arion helps with generative AI digital signage

We build and run the unglamorous half of this: the platform that holds the approved asset, schedules it and proves it played. Arion Signage, at techarion.com/services/arion-signage, is our cloud digital signage product for menu boards, retail and corporate screens, with playlists, dayparting, templates and an approval step, running on Raspberry Pi and Android players so a small network is not a capital project. Because it is API-driven, a generated asset can be pushed from any pipeline and still land behind the same human sign-off. We also do the part a model cannot: real product photography and video at techarion.com/services/product-photography, which is where the reference images that make generated creative look expensive come from. If your screens are advertising inventory rather than an in-store channel, techarion.com/services/dooh-advertising covers the network side. We will tell you plainly which assets are worth generating and which still need a camera.

Frequently asked questions about generative AI digital signage

What retail and restaurant operators in Hyderabad, Chennai and Bengaluru ask before turning generation on.