AI Product Label Test: 9 Attempts Across Ideogram, Krea and Grok
Three product-label prompts, six returned images, and three provider blocks. Every first attempt is shown.
On September 7, 2026, we submitted three fixed fictional product-label prompts to each model through Fal. Every first attempt is retained below, including all three blocks. No images were rerolled, repaired, or replaced.
| Model route | Attempts | Images returned | Provider blocks |
|---|---|---|---|
| Ideogram 4 | 3 | 0 | 3 |
| Krea 2 Medium | 3 | 3 | 0 |
| Grok Imagine | 3 | 3 | 0 |
How the test was run
Each model received the same visible prompt within a row and a 4:3 landscape request. Ideogram used BALANCED rendering with prompt expansion set to None. Krea used medium creativity. Grok used 1k resolution. Ideogram and Krea received seed 20260907; Grok exposes no seed in this route. The seed is not a shared random state across models. Resolution and internal model processing differ.
This is text-to-image generation of fictional packaging. No reference packshot, brand artwork, or real product was supplied. It therefore tests whether a workflow invents the requested label and composition, not whether it preserves an existing SKU.
Exact-text checking rule
We inspected every returned image against the required wording. A complete label passes only when every requested character can be visually verified, including capitalization, punctuation, numerals, and units. Wrapping is noted separately. Ambiguous small characters are marked not verified rather than guessed. These are AI-assisted editorial visual checks, not OCR scores, blinded human-panel results, or reliability estimates.
Short two-line label
Prompt and test settings
A premium commercial product photograph of exactly one upright frosted glass serum bottle on a pale limestone slab beside a shallow reflective pool. The bottle has a flat front-facing cream label. Print exactly these lines on that label, top to bottom: "MORNING DEW", "VITAMIN C SERUM". Preserve every character, number, symbol and capitalization; do not add any other text anywhere. The whole bottle and cap are visible. Warm sunrise light from the left, visible condensation droplets, restrained cream and sage palette. Place the bottle in the left half and leave the right third of the image empty for later ad copy. Landscape 4:3 composition. No other bottles, no watermark.
- Ideogram 4: Fal returned HTTP 422 content_policy_violation with a generated-image safety-filter message. No image was supplied. Cause not independently diagnosed.
- Krea 2 Medium: The brand is readable, but the smaller serum line is distorted. An unrequested forearm enters the lower-left corner.
- Grok Imagine: Both requested lines are readable. The whole bottle is visible in the left half with open space to its right.
Punctuation, numerals and units
Prompt and test settings
A premium commercial product photograph of exactly one upright frosted glass serum bottle on a pale limestone slab beside a shallow reflective pool. The bottle has a flat front-facing cream label. Print exactly these lines on that label, top to bottom: "FIELD & FORM", "NIACINAMIDE 10% + ZINC 1%", "30 mL / 1.0 fl oz". Preserve every character, number, symbol and capitalization; do not add any other text anywhere. The whole bottle and cap are visible. Warm sunrise light from the left, visible condensation droplets, restrained cream and sage palette. Place the bottle in the left half and leave the right third of the image empty for later ad copy. Landscape 4:3 composition. No other bottles, no watermark.
- Ideogram 4: Fal returned HTTP 422 content_policy_violation with a generated-image safety-filter message. No image was supplied. Cause not independently diagnosed.
- Krea 2 Medium: The brand and ingredient names are readable. The volume line does not clearly retain the requested decimal in 1.0; the ingredient line also wraps.
- Grok Imagine: The brand and ingredient wording are readable. The small volume-unit line needs closer inspection; it is not counted as verified exact text.
Long ingredient line and batch code
Prompt and test settings
A premium commercial product photograph of exactly one upright frosted glass serum bottle on a pale limestone slab beside a shallow reflective pool. The bottle has a flat front-facing cream label. Print exactly these lines on that label, top to bottom: "NORTH COAST LABORATORY", "CERAMIDE + PEPTIDE BARRIER SERUM", "FRAGRANCE-FREE | BATCH 07-A". Preserve every character, number, symbol and capitalization; do not add any other text anywhere. The whole bottle and cap are visible. Warm sunrise light from the left, visible condensation droplets, restrained cream and sage palette. Place the bottle in the left half and leave the right third of the image empty for later ad copy. Landscape 4:3 composition. No other bottles, no watermark.
- Ideogram 4: Fal returned HTTP 422 content_policy_violation with a generated-image safety-filter message. No image was supplied. Cause not independently diagnosed.
- Krea 2 Medium: The brand is readable. The bottom fragrance and batch line is visibly distorted, so the complete label fails the exact-text check.
- Grok Imagine: The brand and ingredient wording are readable. The batch-code characters are not clear enough to verify every character; the complete label is marked uncertain.
What to check before using a generated product ad
- Verify the label at full sizeRead every line character by character. Keep unknown characters unresolved instead of inferring the intended word.
- Check the final ad cropA line that is correct at full resolution can become unreadable after cropping or resizing. Inspect the size customers will actually see.
- Check composition separatelyLook for the requested bottle position, cap, pool, lighting and room for copy. Watch for unrequested objects, such as the forearm in the short Krea output.
- Preserve real packaging deliberatelyFor a real product, use approved artwork and a packshot reference. This fictional-label test provides no evidence of actual packaging fidelity.
Limits and the earlier comparison
Three hand-written prompts and one first attempt per model and prompt cannot measure repeated-run reliability. The three Ideogram errors report safety filtering at the provider route; this run does not identify why those benign product prompts were blocked or show that other Ideogram access paths behave the same way. No safety setting was disabled and no blocked job was resubmitted.
The July five-task comparison includes an earlier serum prompt with visible Ideogram output. It used Magic Prompt expansion and a different prompt, so it is separate evidence and is not substituted for a blocked September attempt.