Donatello AI Benchmark / Protocol & evidence
Eight reference tasks with separate checks for fidelity, identity, style and artifacts. No reference-quality ranking is available yet.
Initial release: protocols and documentation estimates only. No comparative generation results, public votes or quality winners have been published.
8 cases × 3 planned attempts per model. No models have been enrolled and no attempts have been executed. For two models, the full eight-case reference set would require 48 attempts; this is a planning count, not a cost quote or observed sample.
| Frame / target | 1:1 · 1024 × 1024; disclose unsupported sizes |
|---|---|
| Attempt allocation | Three registered attempts per prompt and model, including technical failures |
| Enhancement | Off where exposed; otherwise record provider default and place in a separate comparison stratum. |
| Outputs / winner | Not measured / no winner |
| Price / latency / failure rate | Not measured—not zero |
Three cases reuse the same public synthetic vase. The five remaining cases must not run until their cleared references are registered. A face-preservation benchmark cannot be substituted with a vase test.
Previously published first-party synthetic ceramic-vase image. Reused only as a disclosed benchmark input, never as a benchmark output.
A Donatello-origin reference may introduce source-domain bias. Add a rights-cleared independent source stratum before generalizing reference results. No real person's likeness is used.
SHA-256: 0089dad7f8cb5eaee9e90c797d28e7a0919214602311b5a13af649253fb685bc · 41,080 bytes
Original provenance record →Use the text exactly as published. The checks establish minimum task adherence; they are not an aesthetic score. A protocol amendment must not be optimized selectively after seeing one platform's output.
8 of 8 prompts · No outputs or evaluations recorded.
Place the same fictional adult character from the reference in a sunlit bookshop, preserving the character's facial proportions and hairstyle.
Required checks: Identity · Hair and proportions · Bookshop setting
Reference IDs: None supplied. Planned attempts per model: 3. Winner: not measured.
Place the exact blue ceramic vase from the reference on a dark wooden shelf, preserving its shape, opening and glaze color, with soft side light.
Required checks: Vase silhouette and opening · Blue glaze identity · New shelf setting
Reference IDs: blue-vase. Planned attempts per model: 3. Winner: not measured.
Show the same fictional adult wearing the exact jacket from the supplied references, full-length front view on a plain grey background. Preserve the garment's cut and fastening details.
Required checks: Character identity · Garment details · Full-length front view
Reference IDs: None supplied. Planned attempts per model: 3. Winner: not measured.
Change only the background of the consented adult portrait to a plain light blue wall. Preserve the face, expression, hairstyle and clothing.
Required checks: Face unchanged · Expression and clothing unchanged · Only background edited
Reference IDs: None supplied. Planned attempts per model: 3. Winner: not measured.
Create a small ceramic teapot using the glaze color, material finish and soft-daylight photographic style of the supplied vase reference. Do not reproduce the vase shape.
Required checks: Glaze and photographic style · Recognizable teapot · Not a copied vase
Reference IDs: blue-vase. Planned attempts per model: 3. Winner: not measured.
Place the product from reference one on the table from reference two, with the lighting from reference three. Preserve the product identity and table material.
Required checks: Product identity · Table material · Third-reference lighting
Reference IDs: None supplied. Planned attempts per model: 3. Winner: not measured.
Keep the same blue vase from the reference, rotate the viewpoint slightly to the right and add one white flower inside. Preserve the rim, body proportions and blue glaze.
Required checks: Vase consistency · Slight viewpoint change · Exactly one white flower
Reference IDs: blue-vase. Planned attempts per model: 3. Winner: not measured.
Move the fictional branded package from the reference onto a white pedestal. Preserve every letter, symbol and label placement exactly; change only the setting.
Required checks: Exact lettering · Label layout · White pedestal
Reference IDs: None supplied. Planned attempts per model: 3. Winner: not measured.
Review the full methodology before executing. Input readiness does not establish provider compatibility, permission to publish outputs or an approved budget.
Explore Donatello's image generator →