Real PosterSpeak workflow

From campaign copy to a usable poster asset

BACK TO SCHOOL SALE

Save 20% on selected school essentials

August 1–15

Shop now

We generated three posters, checked them at phone size and kept Candidates 1 and 2. Each selected poster became a separate 8.5-second video; the images were not combined into a slideshow.

Model
GPT Image 2
Credits
21 total: 15 poster + 2 voice + 4 for two video exports
Selected back-to-school poster candidate with blue backpack, notebooks and sale information
Poster candidates

Three directions from the same copy

Energetic blue and yellow back-to-school sale poster candidate

Candidate 1

Selected for video

The brush-stroke headline feels quickest and loudest in a feed. We kept it as the more energetic test variant.

Dark blue back-to-school sale poster candidate with a large backpack

Candidate 2

Selected for video

This was the easiest version to scan on a phone: sale first, discount second, dates third. It became video version 1.

Bright yellow back-to-school sale poster candidate with stationery

Candidate 3

Not exported

The yellow background stands out, but the layout reads more like a catalogue cover. We left it as a static option and did not spend credits exporting it.

Final delivery

Two independent video versions

Ready to publish

Video version 1

A separate export made from Candidate 2, using the same campaign copy and voiceover.

Format
MP4
Ratio
9:16
Length
8.5 seconds
Ready to publish

Video version 2

A separate export made from Candidate 1, using the same campaign copy and voiceover.

Format
MP4
Ratio
9:16
Length
8.5 seconds

The input was only four lines

We used a headline, one discount, a product category, dates and the call to action “Shop now.” Nothing else was added. That kept the offer readable on a phone and gave the image model fewer chances to invent unsupported details.

The campaign was already in English, so translation stayed off. “Shop now” was entered as the closing line so it could appear on both the poster and the voice track.

We directed the look, not the sales message

The visual note asked for modern school supplies, bold retail typography, high contrast and enough breathing room around the top and bottom interface areas. We also told the model not to add prices, brand names, contact details or extra claims.

That distinction was useful in practice: the words stayed fixed while the model varied the hierarchy, colour balance and product arrangement.

Why we kept two of the three posters

Generating three candidates with GPT Image 2 cost 15 credits. We opened each at phone size before choosing. Candidate 2 had the cleanest reading order; Candidate 1 was busier but more energetic. Together they made a sensible A/B pair.

Candidate 3 was not broken. It simply looked closer to a catalogue cover, so exporting it would have spent two more credits without adding a meaningfully different ad angle.

The final output was two separate videos

Each selected poster received its own MP4. Both files are 9:16, run for about 8.5 seconds and use the same voiceover, so an editor or media buyer can compare the visual treatment without changing the message.

The bill was 21 credits: 15 for the three poster candidates, 2 for the voiceover and 4 for two video exports. The unselected poster did not incur a video charge.

Reusable back-to-school copy template

Replace the brackets with verified campaign details. If you have another benefit to test, make it a second creative rather than squeezing another line into the same poster.

  • Headline: BACK TO SCHOOL [SALE OR EVENT]
  • Offer: Save [DISCOUNT] on [PRODUCT CATEGORY]
  • Dates: [START DATE]–[END DATE]
  • Closing line: [SHOP NOW OR LEARN MORE]