Creator Tool Lab

A real prompt-to-image run, tool checks with disclosed limits, and calculations you can reproduce.

Published October 24, 2026 · Updated November 10, 2026 · TrendPromptLab Editorial / ALL IN ONE

We ran a prompt assembled by our product photo builder through ChatGPT’s image-generation tool on October 24, 2026. We also captured the actual output of ten text builders and checked three planning calculations. The briefs and financial assumptions are fictional; the rendered image and captured builder outputs are real artifacts from these runs. This is a bounded demonstration, not a cross-model benchmark or evidence of customer results.

1. From product brief to a generated image

The inputs were a matte black insulated travel mug, clean commercial style, warm-gray studio, softbox with rim light, three-quarter angle, and square ecommerce composition. No reference photograph was supplied. We used the exact builder output below without adding a product specification.

Exact prompt sent to the image-generation tool

Create an original clean commercial product photograph of a matte black insulated travel mug, styled in a seamless warm-gray studio. Use large softbox lighting with a controlled rim light and a three-quarter hero angle at product height. Compose for square ecommerce listing with generous negative space. Accurate materials, realistic reflections and shadows, clean edges, believable scale, commercially polished color, no people unless requested, no extra objects competing with the product, no distorted packaging, no invented logos, no watermark, no readable text. Keep the product as the unmistakable focal point.

Generated matte black mug with a handle, silver rim and transparent lid on a warm-gray studio background — Actual generated result. Synthetic concept image; no real SKU or physical product was tested. One run, no image edits. No model version or seed was recorded, so this case does not support a claim about a specific model’s performance.

Our inspection of this output

Requested	Observed	Decision
Matte black mug	Dark textured body with a black handle	Color and surface direction are visible; handle design was invented.
Warm-gray studio	Plain gray-beige surface and backdrop with a cast shadow	Suitable concept backdrop; not proof of a physical studio shoot.
One product as focal point	One mug fills most of the square frame; no extra props or visible writing	Readable silhouette, but less negative space than requested.
No reference SKU	A clear lid, silver rim and black tab were added	Reject as a real sales photograph unless these details match actual stock.

The practical correction

For concept art, this may be usable after visual review. For an existing product listing, supply an authorized reference photograph and enumerate the lid, handle, rim and dimensions that must stay unchanged. Compare the output with the item that will ship. A polished result can still be the wrong product.

A possible next prompt addition is: “Use the supplied reference as the product specification. Preserve its exact handle, lid, rim, proportions and surface finish. Change only the studio background and light.” That revision is a proposed next test; it was not run here.

Load the same inputs in the product photo builder →

From the generated concept to a finished thumbnail

We loaded the same mug into our local thumbnail studio, added the headline “A POLISHED IMAGE. THE WRONG PRODUCT?” and exported the PNG below at 1280 × 720. The headline describes the documented inspection, rather than inventing a performance result. The editor’s planning guides are absent from this actual downloaded file.

Actual exported thumbnail with the headline A polished image. The wrong product? beside the generated mug — Actual browser-exported layout, October 24, 2026. Synthetic source image; illustrative editorial use.

Create and download your own thumbnail →

Measured layout case: what a square photograph loses in a split thumbnail

Question: Can the whole source photograph fit into the right half of a 1280 × 720 thumbnail without changing the image? On October 24, 2026, we loaded our existing 1254 × 1254 synthetic mug image into the browser editor and exported two PNGs. We kept the headline, subline, colors, photo side and centered position fixed. Only the framing mode changed. These are actual browser exports, not simulated screenshots or a new image-model run.

Actual fill-mode export with the square mug image filling the right half of the thumbnail — Fill mode: the photograph fills all 640 × 720 pixels of its panel; 88.9% of source area remains.

Actual full-photo export showing the whole square image and dark margins above and below it — Full-photo mode: all source pixels remain in view; the photo occupies 88.9% of the panel, with 40-pixel margins above and below.

Held constant	Fill photo area	Fit full photo
Source and photo panel	1254 × 1254; 640 × 720	Same source and panel
Rendered photograph	720 × 720	640 × 640
Clipping at centered position	40 rendered pixels at each horizontal edge; about 69.7 source pixels per edge	None
Source area retained	88.9%	100.0%
Photo-panel area occupied	100.0%	88.9%

How to reproduce the numbers

For fill mode, scale = max(640 ÷ 1254, 720 ÷ 1254), so the source becomes 720 × 720. The 640-pixel panel removes 80 rendered pixels of width. A centered position splits that loss into 40 pixels at each edge. Source-area retention is (640 ÷ 720) × (720 ÷ 720) = 88.888…%. For full-photo mode, scale = min(640 ÷ 1254, 720 ÷ 1254), giving a 640 × 640 photograph and 80 pixels of unused panel height.

Our editorial decision

The mug silhouette, lid, rim and handle remain visible in both of these centered exports; the fill version removes background at the horizontal edges. Therefore 11.1% source-area loss does not mean 11.1% of the product was removed. A large retention percentage also cannot certify that a small label or accessory survived. Use the full-photo version when the entire source matters. Use the position sliders when a cropped version needs to protect an off-center detail, then inspect the actual subject and the 320-pixel preview.

The two files demonstrate a geometric tradeoff, not which thumbnail viewers prefer. We did not measure clicks, readability with participants or audience retention. Browser fonts and image resampling can vary; the case files preserve the actual exports from this run.

Download exact settings and arithmetic (.txt) · Download the synthetic source image · Reproduce the crop in the editor →

2. Pricing: the denominator matters

Use 20 build hours × $40/hour + $100 fixed cost = $900 build cost. Add $50 of overhead and divide by 50 target sales: $19 to recover per sale. With a 10% fee and 20% reserve, only 70% of price remains for recovery and target profit.

Break-even = 19 / (1 − 0.10 − 0.20) = $27.14 (display rounded)
With $20 target profit per sale = (19 + 20) / 0.70 = $55.71
At 25 target sales: recovery = 950 / 25 = $38; break-even = $54.29
At 50% fees + 50% reserve: no finite break-even price.

We corrected the calculator so a combined 100% deduction produces a clear invalid-input message instead of an arbitrary price. Fractional target sales and empty numbers are also rejected. Currency formatting now shows cents where relevant. The reserve is a planning input, not a tax-rate recommendation.

Reproduce the pricing calculation →

3. Revenue: a positive gross can still leave almost nothing

With 5,000 visitors, a 3% subscriber rate and a 2% purchase rate, the scenario has 150 subscribers and 3 purchases. At $19, gross is $57. A 5% refund assumption removes $2.85. A 5% + $0.50 payment fee on each original sale costs $4.35. After $50 overhead, contribution is −$0.20. The break-even count is 4 sales because each sale contributes $16.60 before overhead.

This model assumes processing fees are charged on original purchases and not returned with refunds. Fractional purchases in smaller scenarios are expected values, not actual partial orders. It excludes labor, taxes and chargebacks. Copy and download now recalculate from the latest inputs, so changing a field does not export old results.

Reproduce the revenue scenario →

4. Subscriptions: find overlap before forecasting income

Enter illustrative monthly assistant costs of $20 and $20, a $15 design tool and a $50 budget. The total is $55/month or $660/year, $5 over budget. The assistant category is flagged twice. These are example inputs, not current provider prices; they do not establish that either subscription is unnecessary. Compare the actual task and cancellation terms first.

Run the stack-cost example →

5. Inspect all ten text-builder outputs

Each linked page exposes the complete input settings, captured output, a button to load the example and a task-specific editorial critique. These are rule-based drafts. No output is claimed to improve rankings, views, sales or image quality without a separate measured test.

6. Input-variation audit: 211 cases across ten tools

We ran each of the ten text builders with available alternative selections, blank inputs, and multilingual or special-character text. Across 211 input combinations, the current source produced an output without a runtime failure, an empty result, or the literal word “undefined.” The combinations covered input handling; they were not 211 independent users or a quality score.

Audit item	What we checked	What it cannot show
Scope	Ten text builders; alternate select values, empty fields, multilingual and special-character entries	How often real visitors use these inputs
Result	211 combinations; no runtime failure, missing output or “undefined” result recorded	Whether every draft is accurate, useful or persuasive
Human review	Separate page examples include full settings, captured output and task-specific critique	Whether a viewer prefers the wording or achieves a business outcome

Two defects found and corrected

Spoken-word estimate: production labels such as “SAY:” were included in a Shorts script's spoken-word count. The counter now excludes those labels and uses the speech calculator's word-count rule.
Copy feedback: one prompt page could report a successful copy after both clipboard paths failed. It now surfaces the failure and leaves the text available for manual selection.

These checks test software behavior and visible error reporting. They do not judge every generated line, test an external image model, establish user engagement, or predict an advertising review.

What this review does not establish

One image is not an image-model evaluation. Arithmetic checks do not validate the visitor or sales assumptions. Input-variation checks do not prove semantic quality or audience value. A working template does not replace filming, editing, source checking or audience research. We report the inputs, environment, observations and limits of each test.