What changed

Runware lists Pruna’s P-Video-2-Pro with a 17 September 2026 date. Its documentation describes text-to-video and image-to-video generation, an optional final reference frame, generated audio, and clips lasting 5 to 15 seconds. Listed outputs are 480p or 768p at 24 frames per second. Those are documented capabilities, not an independent quality test.

The integration uses asynchronous delivery: submit a generation task, retain its identifier, then retrieve the result when it completes. A workflow that expects an immediate finished video needs to account for that delay.

The useful question is what happens after generation

A finished file is the beginning of a production decision. Does the character still look right? Can the shot cut against the previous one? Is the audio usable? A technically successful request can still produce material that an editor rejects.

For a small automated video workflow, separate three states: requested, delivered and accepted. Keep the source prompt and reference alongside each result. That makes it possible to compare an accepted clip against the actual request, instead of judging a provider from its best isolated example. This is a proposed workflow design, not a claim about this model’s performance.

A small test before a large integration

Choose one reference image you own and one detail that must remain consistent. Request a short three-shot sequence with the same subject. Use the final frame of each accepted clip to inform the next shot where the API permits it. Keep the framing and output settings comparable across the tools being evaluated.

Record the time until each file is playable, the billed amount, whether the key detail survived, and how many seconds you would actually use. Include rejected generations in the calculation. The measure that matters for this test is total spend divided by accepted seconds, alongside continuity and editorial effort.

Do not scale the workflow until you can explain its failure cases. A retry policy, an approval step and a way to recover a completed job are worth deciding before connecting the model to client delivery. No paid test was performed for this article; speed, visual continuity and savings remain open questions.

Sources & verification

Prepared with AI assistance from the linked primary documentation. The evaluation process is editorial guidance. No model benchmark or paid generation was performed.