What happened in six attempts?
ERNIE-Image Turbo completed 5 of 6 attempts. Three of the 5 completed posters reproduced all three specified fields correctly. Dates and prices matched in 5/5 images each; venue fields matched in 3/5. The other two images changed the Chinese venue label 地点 to 地价 or 地期. The actual venue name remained correct. One request failed with HTTP 429 rate limiting and was not retried.
Across the five completed images, 13/15 fields were visually exact, 2/15 mismatched, and 0/15 were uncertain. These are counts for one fictional workshop, two equivalent prompts and one size. They are not a general accuracy estimate for the model.
Both editions report the same run on the hosted ERNIE-Image Turbo generator.
Compare the original images
The six-panel overview keeps the failed attempt in its original position. The detail comparison uses the same pixel rectangle for every successful image, enlarged with nearest-neighbor sampling; it does not repair text. Open the full-size figures or each unchanged original below.
ERNIE-Image Turbo: 6 attempts, 5 outputs; successful images at original 512px width; failed jobs have no replacement image.
ERNIE-Image Turbo: comparison crop [60, 280, 490, 435]; nearest-neighbor 2× enlargement; no text repairs.
The text we asked for
The fictional workshop is “周末创作 / WEEKEND CREATE”. These are the three complete reference strings, including labels, capitalization and punctuation:
| Field | Exact reference text |
|---|---|
| Date | 日期 DATE: 2026-12-05 |
| Price | 票价 PRICE: ¥29.90 |
| Venue | 地点 VENUE: 云杉工坊 / SPRUCE STUDIO |
Method and scoring
Test date: 2026-11-13. Both language versions report the same image set.
We submitted A-41731, B-41731, A-41732, B-41732, A-41733, B-41733 in that order through the normal site interface. A and B request the same content; B expresses the information as explicit rows. The UI and request records carried seeds 41731, 41732 and 41733. The failure at A-41732 left only two seed pairs with both outputs. This verifies requested seeds, not identical backend execution or a pinned checkpoint.
We used 512 × 512 pixels, one image, and the interface’s free/0-credit tier. The six-attempt cap included failures; we did not replace the failed A-41732 request. After its rate-limit response we waited before submitting the remaining, different preregistered items. For every attempt we retained the date, model label, selected size, image count, prompt, requested seed and available output. Backend checkpoint revision, inference settings, and hardware remain unknown unless separately disclosed by the service; this is not a local checkpoint reproduction.
For every completed output, transcribe date, price and venue literally. A field passes only when the entire line matches its reference, including Chinese characters, English case, digits, the decimal point, currency symbol, colon, hyphens and slash. We judge visible text, not Unicode code points recovered from raster pixels. Record line breaks and read a wrapped field in order, replacing each line break with one space for comparison; a wrap alone does not fail the field. Actual ambiguous characters or separators remain uncertain. A clearly visible error is a mismatch even if another part is unreadable. Do not silently correct characters. Report exact fields over scoreable fields, all-three-field posters over scoreable posters, and failed jobs separately. Two AI reviewers inspected the five originals independently, then compared their transcriptions. They agreed on all 15 field readings and statuses, with no unresolved disagreement. Expected strings had been read during preparation, so this was not a fully blind review.
All attempts and field results
| Attempt | Job | Date | Price | Venue |
|---|---|---|---|---|
| A-41731 | completed | exact | exact | mismatch |
| B-41731 | completed | exact | exact | exact |
| A-41732 | HTTP 429 | not scored | not scored | not scored |
| B-41732 | completed | exact | exact | exact |
| A-41733 | completed | exact | exact | mismatch |
| B-41733 | completed | exact | exact | exact |
B produced three all-field-correct images out of three completed outputs; A produced zero out of two, plus the failed request. With only two complete pairs and one missing output, this case does not establish that B is generally better.
Literal field transcriptions
Each line below is the observed text, not a corrected version. Compare it with the reference table above.
A-41731
- Date:
日期 DATE: 2026-12-05— exact - Price:
票价 PRICE: ¥29.90— exact - Venue:
地价 VENUE: 云杉工坊 / SPRUCE STUDIO— mismatch
B-41731
- Date:
日期 DATE: 2026-12-05— exact - Price:
票价 PRICE: ¥29.90— exact - Venue:
地点 VENUE: 云杉工坊 / SPRUCE STUDIO— exact
B-41732
- Date:
日期 DATE: 2026-12-05— exact - Price:
票价 PRICE: ¥29.90— exact - Venue:
地点 VENUE: 云杉工坊 / SPRUCE STUDIO— exact
A-41733
- Date:
日期 DATE: 2026-12-05— exact - Price:
票价 PRICE: ¥29.90— exact - Venue:
地期 VENUE: 云杉工坊 / SPRUCE STUDIO— mismatch
B-41733
- Date:
日期 DATE: 2026-12-05— exact - Price:
票价 PRICE: ¥29.90— exact - Venue:
地点 VENUE: 云杉工坊 / SPRUCE STUDIO— exact
The two errors are in the second character of the Chinese venue label: 点 → 价 in A-41731 and 点 → 期 in A-41733. The date, ¥29.90, and 云杉工坊 / SPRUCE STUDIO are readable and match their reference text in all five completed images. A-41732 has no transcription because no output was produced.
Download the five unmodified originals
- A-41731 — unchanged original · 512 × 512 · WebP · SHA-256
69c9296f694ff1d1e2bb902f4135416a4ee07c3dd4e0586573bd4cffbaf1b094 - B-41731 — unchanged original · 512 × 512 · WebP · SHA-256
6b19af215644eee9e371c6eda920f90e760b18bcb61fae0e6b26450f6879842e - B-41732 — unchanged original · 512 × 512 · WebP · SHA-256
7acd9bbf874b9ff93ecb23eb0a40b6ba8e8729336cf47d4c0f8460a61dda4281 - A-41733 — unchanged original · 512 × 512 · WebP · SHA-256
f9aeb2ad102481d72cfc135f2a42e599192dd32927c7ebb244eab6a4c4ab702b - B-41733 — unchanged original · 512 × 512 · WebP · SHA-256
f7779b73de6c9101762dbb8034dd5d200085a8759d965a7f7db6d7db4ec9479c
Original file links preserve the actual output bytes. A-41732 has no image because the request failed with HTTP 429; no replacement is shown.
Reuse the two prompts
Both prompts below are the actual English instructions for the same bilingual output. Keep the quoted strings unchanged when repeating this case.
Prompt A
Create a bilingual Chinese-English poster for a fictional community art workshop titled "周末创作 / WEEKEND CREATE". Use a clean blue and white design with a large title at the top and three separate, clearly readable lines below. Print these lines exactly: "日期 DATE: 2026-12-05"; "票价 PRICE: ¥29.90"; "地点 VENUE: 云杉工坊 / SPRUCE STUDIO". Do not add other words. Keep every digit, decimal point, slash and currency symbol unchanged.
Prompt B
Design a blue-and-white bilingual poster for the fictional community art workshop "周末创作 / WEEKEND CREATE". Layout: top, large workshop title; below, three distinct information rows. Row 1 must read "日期 DATE: 2026-12-05". Row 2 must read "票价 PRICE: ¥29.90". Row 3 must read "地点 VENUE: 云杉工坊 / SPRUCE STUDIO". Render the quoted wording literally, preserving all digits and punctuation. Include no additional text.
Repeat and proofread your own poster
- Open the image generator, check that the selected model is ERNIE-Image Turbo, and record the settings currently available. The free allowance and model menu may change.
- Select the same recorded dimensions and one image. Paste one full prompt, set the seed if available, and save the original output before editing it. Respect the service’s current quota; these results do not promise free or unlimited retries.
- Compare each full text line against the reference at a readable zoom. Check
¥29.90character by character; a missing decimal point changes the price. - Keep every attempt, including failures. For an actual event, correct the final typography in an editor and have another person verify dates, prices and addresses before publishing.
Limits of the evidence
One fictional event and at most six attempts cannot establish general accuracy, a reliable advantage for one wording, or performance in other sizes, languages or model versions. The two prompt variants also cannot isolate every layout or wording effect. This test scores three specified fields, not aesthetics, the workshop title, accessibility, print quality or all extra text. Any extra text should still be inspected before using a poster. Hosting settings may change, and equal seed labels alone cannot guarantee an identical generation environment.
Sources and related reading
The findings in this article come from this experiment’s original outputs and field transcriptions. Baidu’s official ERNIE-Image-Turbo model card identifies the model family; its examples and benchmarks are not substituted for our results. For background on layout and multilingual text, see the earlier structured-layout guide and text-rendering guide. Their general advice is separate from this small exact-text experiment.

