Pruna AI models, access and benchmark guide
Pruna AI model and access guide with a practical US-English evaluation brief and links to publisher materials.

👁 111 reads
Reference footage can change the economics of a Pruna job. Its reference model distinguishes the material supplied to the request and the video it emits. I would keep resource IDs for the source and delivery separately before ordering variations.
A reference-led P-Video-2-Pro order can spend money on two lengths of footage: the clip supplied and the material returned. In a shop-window loop, a three-second reference reused across several trials can therefore affect the quote repeatedly. Build the shot list with a separate reference-duration column, write down the selected mode and record rejected versions. The benchmark can inform which model gets the first trial; only the chosen provider and settings establish the price of that particular loop.
Choosing the first job
Animate an original clock face while keeping its numbers fixed; compare the reference length with the delivered duration.
Models in the November 17 snapshot
| Model and track | Placement | Reported score |
|---|---|---|
| P-Video-2-Pro (AA text to video) | 13 | 1128 Elo |
| P-Video-2-Pro (AA image to video) | 8 | 1117 Elo |
Release record. Publisher documentation identifies the product family. The ranked configurations include P-Video-2-Pro.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- https://docs.api.pruna.ai/guides/models (accessed November 17, 2026)
- https://huggingface.co/PrunaAI (accessed November 17, 2026)
- https://artificialanalysis.ai/video/leaderboard/text-to-video (accessed November 17, 2026)
- https://artificialanalysis.ai/video/leaderboard/image-to-video (accessed November 17, 2026)
More on this from QUANTSLANT
benchmarks·PixVerse models, access and benchmark guidePixVerse model and access guide with a practical US-English evaluation brief and links to publisher materials.