P-Video-2-Pro-Reference is Pruna's MiniMax H3-based model for generating new footage from a prompt and image or video references. It adds a distinct access and pricing choice to the H3 ecosystem. A reference can guide appearance or motion; it does not turn every generation into an exact edit of existing footage.
The provider's model page, usage guide, and API contract were first checked October 16, 2026; the developer contract was rechecked October 22. Examples below explain those documents and arithmetic, rather than paid predictions. H3Vid has not integrated this provider model.
Reference is a different model identity
Use Model: p-video-2-pro-reference on Pruna's predictions API. The ordinary p-video-2-pro has a different input contract. The Reference variant requires a prompt with at least one image or video reference; text alone is rejected. Audio is optional and is not a substitute for that visual input.
The ordinary Pro documentation is useful for comparison: it describes text and optional endpoint stills, and a quality mode. Do not copy those capabilities into a Reference request. Reference offers cost and speed, with speed the default. It has no quality or draft mode. Pruna says cost preserves speed-mode output quality; that is a provider claim, not our benchmark.
Assign a role to each reference
Start with one clear still and a short action. A product image can define the object; a separate video can communicate motion. Write explicitly which details each input should contribute, especially if its background should be ignored. Adding more media without defining their roles can make a request harder to diagnose.
An original example brief is: “Use the still for the red enamel mug and white rim. Use the clip only for a slow clockwise camera movement. Keep the mug on a wooden table at dawn. Gentle steam, quiet wind, no speech.” This is a suggested prompt structure, not a tested result or guaranteed preservation rule.
The model supports up to nine images including first and last frames, three videos, and three audio files. The developer contract limits combined video and audio durations to fifteen seconds per category, with additional file-size and format limits. Check these before uploading a large library. Start and end stills consume image slots rather than adding two slots beyond nine.
Soft frames and kept audio
first_frame_image and last_frame_image guide the opening and closing. They are soft constraints, so inspect the result when an exact frame matters. The developer portal says endpoint stills trigger turbo prompt upsampling even when prompt_upsampler is off. Keep the submitted prompt and actual settings when investigating unexpected rewriting.
Audio normally serves as a reference. keep_audio requires exactly one audio file and changes how it is used: the uploaded track becomes the video's audio. Short tracks are followed by generated sound; long tracks are cut. A track within one second of the requested duration can determine length. Listen across the entire result rather than assuming the requested integer gives the final soundtrack boundary.
The current developer contract adds 1080p, 2K and 4K profiles by upscaling the 768p output with p-video-upscale, keeping 24 fps. Native profiles remain 480p and 768p, with requests from five to fifteen seconds. Choose adaptive aspect ratio when the composition should follow the references. Resolution and mode are separate choices, so record both when comparing cost or execution time.
Calculate the whole reference bill
The entry rate on the marketing page covers output seconds. Inputs also incur charges. At the checked rates, 480p cost mode charges $0.015 per output second, $0.005 per image, $0.03 per reference-video second, and $0.0015 per reference-audio second. The corresponding 768p output rate is $0.025; its reference-video rate is $0.06. Speed-mode rates differ.
For five billed output seconds at 480p in cost mode with one ordinary image, the arithmetic is $0.075 + $0.005 = $0.080. Adding five billed seconds of reference video adds $0.15. Adding five billed audio seconds adds $0.0075. The resulting illustrative total is $0.2375. This uses stated billed quantities, excludes unrelated services, and is not an invoice or H3Vid credit quote.
The docs themselves contain a tension: they label $0.075 with no references as a cheapest request while also requiring a visual reference. Treat that number as output-only arithmetic, not an actionable total for an accepted reference request.
The developer portal further explains output billing as int(num_frames / 24) and caps each reference clip at output length plus half a second, within the combined cap. Thus requested duration, actual media duration, and billed seconds can differ. Log final usage instead of multiplying the input duration blindly.
Higher-resolution delivery and final usage
At the October 22 check, upscaling adds $0.02 per output second for 1080p, $0.04 for 2K and $0.08 for 4K to the selected 768p mode's rate. Thus cost-mode output rates become $0.045, $0.065 and $0.105 per second; speed-mode rates become $0.055, $0.075 and $0.115. These are enhanced delivery profiles, not evidence of native high-resolution generation.
Reference video still uses the 768p reference rate for those upscaled outputs. Image and audio rates continue to depend on mode rather than output resolution. For five billed output seconds at 1080p in cost mode with one image and no other references, the documented arithmetic is $0.225 + $0.005 = $0.230. Confirm finished usage and rounding against the returned run, especially when kept audio changes the output length.
The current documentation's whole-second convention means a fifteen-second request can encode 345 frames and bill fourteen output seconds. That example belongs to Pruna's wrapper; it is not a universal H3 frame rule. We have not verified an invoice or the failed-run refund behavior, so this update does not claim a fixed refund policy.
Make one reviewable API experiment
Follow Pruna's quickstart to obtain access and upload one file. Submit an asynchronous prediction with the exact Reference model header, the returned image URL, an explicit mode, and a short requested duration. Keep the prediction ID, poll the documented status, then download the completed media. Compare the finished object, motion, sound, and final charge before adding more references.
For an existing H3Vid reference workflow, see Ref2VA. For preserving a source performance with a changed subject, examine the character-swap page. These pages address related creative needs, but neither executes Pruna's named model or supplies its price contract.