Measured / October 6
FLUX.2 [pro] and Ideogram 4.5
Four shared prompts, three repetitions, every output, actual API charges and response times. A technical pilot, not a quality ranking.
Inspect flux.2 [pro] and ideogram 4.5 →Donatello AI Benchmark / Protocol & evidence
Useful evidence. Not a predetermined winner.
A transparent testing programme for AI images, references and real creator workflows. Inspect the protocol, download the inputs and see exactly what has—and has not—been measured.
Protocol v1 and its downloads remain available. The separately registered October 6 pilot adds 24 API attempts, technical results and costs. No public votes, independently reviewed quality scores or quality winners are published.
What is available today?
30Published image prompts · ten categories
8Reference cases · three inputs ready, five pending
24API pilot attempts · no claimed quality winner
01 / Protocol
Image quality
Same prompts, three planned attempts per enrolled model, separate quality, adherence, cost and latency measures.
Inspect image quality →02 / Protocol
Reference fidelity
Inspect source provenance, readiness and checks for identity, products, style and object consistency.
Inspect reference fidelity →03 / Documentation simulation
What does $10 actually buy?
Explore $0, $5, $10 and $20 against three creator briefs. Feature availability is not a completed workflow.
Inspect what does $10 actually buy? →04 / Accountability
Every rule in the open
Sampling, failure accounting, independent review, cost formulas, privacy gates and versioning.
Inspect every rule in the open →Evidence before rankings
The existing gallery proves that particular Donatello outputs were generated. It does not compare equal prompts and attempt counts across providers. The platform comparison pages document features and licences. Neither is relabelled here as a measured quality benchmark.
The four-prompt API pilot is separate from the unexecuted full 30-prompt quality study. Its manifest was frozen locally before generation. Reference testing still needs cleared source assets. A protocol alone cannot answer which tool makes the best image.
Later phases—not live tests
Spanish/English voice, text-to-video, image-to-video, faceless YouTube, music and complete-ad workflows remain future tracks. Native-audio video will be separated from silent video; provider voices from controlled voice; and instrumental music from songs. Public voting and measured leaderboards are not active.