cinemasetfreeprunaai/p-video-2-pro-reference
Pruna's reference-to-video model, built on MiniMax H3. Generate 5–15 second video from a prompt plus up to 9 images, 3 videos, and 3 audio clips, with speed and cost modes.
Capabilities
Cost
Community model (estimated from hardware time)
Input Parameters
| Name | Type | Description | Default | Constraints |
|---|---|---|---|---|
prompt* | string | Text prompt for video generation. | — | — |
aspect_ratio | string | Aspect ratio of the generated video. adaptive uses the most common one among the reference images and videos. | "16:9" | adaptive16:99:164:33:43:22:31:1 |
audio | array | Reference audio files to condition the generation on. Set keep_audio to use a single file as the video's audio. | — | — |
duration | integer | Video duration in seconds. With keep_audio, an audio file within 1 s of it sets the length instead. | 5 | min: 5, max: 15 |
enable_safety_filter | boolean | Check the inputs against the content policy. When off, only the always-on categories (sexual content involving minors, self-harm intent or instructions) are still blocked. | false | — |
first_frame_image | string(uri) | Optional image the video should open on. Combines with the other references. Needs prompt upsampling, so prompt_upsampler=off runs turbo when a start or end frame is given. | — | — |
image | array | Reference images to condition the generation on. | — | — |
keep_audio | boolean | Keep the uploaded audio as the video's audio instead of only using it as a reference. Needs exactly one audio file. The video is generated to match it. If the audio is shorter than duration, the video plays it first and the model generates the rest. If it is longer, it is cut to duration. If it is within 1 s of duration, the video's length follows the audio's. | false | — |
last_frame_image | string(uri) | Optional image the video should end on. Combines with the other references. | — | — |
mode | string | speed runs on 4x B200; cost runs the same model on one B200 (slower, cheaper). | "speed" | speedcost |
prompt_upsampler | string | Rewrite the prompt into a detailed H3 brief that looks at the references first: turbo (our upsampler), max (MiniMax; falls back to turbo), or off. With adaptive aspect ratio, turbo also picks the ratio. | "turbo" | turbomaxoff |
resolution | string | Video resolution. 1080p, 2k and 4k generate at 768p and upscale it (24 fps kept). | "768p" | 768p480p1080p2k4k |
seed | integer | Random seed for reproducible generation. | — | — |
video | array | Reference videos to condition the generation on. | — | — |
promptrequiredstringText prompt for video generation.
aspect_ratiostringAspect ratio of the generated video. adaptive uses the most common one among the reference images and videos.
"16:9"audioarrayReference audio files to condition the generation on. Set keep_audio to use a single file as the video's audio.
durationintegerVideo duration in seconds. With keep_audio, an audio file within 1 s of it sets the length instead.
5min: 5, max: 15enable_safety_filterbooleanCheck the inputs against the content policy. When off, only the always-on categories (sexual content involving minors, self-harm intent or instructions) are still blocked.
falsefirst_frame_imagestringOptional image the video should open on. Combines with the other references. Needs prompt upsampling, so prompt_upsampler=off runs turbo when a start or end frame is given.
imagearrayReference images to condition the generation on.
keep_audiobooleanKeep the uploaded audio as the video's audio instead of only using it as a reference. Needs exactly one audio file. The video is generated to match it. If the audio is shorter than duration, the video plays it first and the model generates the rest. If it is longer, it is cut to duration. If it is within 1 s of duration, the video's length follows the audio's.
falselast_frame_imagestringOptional image the video should end on. Combines with the other references.
modestringspeed runs on 4x B200; cost runs the same model on one B200 (slower, cheaper).
"speed"prompt_upsamplerstringRewrite the prompt into a detailed H3 brief that looks at the references first: turbo (our upsampler), max (MiniMax; falls back to turbo), or off. With adaptive aspect ratio, turbo also picks the ratio.
"turbo"resolutionstringVideo resolution. 1080p, 2k and 4k generate at 768p and upscale it (24 fps kept).
"768p"seedintegerRandom seed for reproducible generation.
videoarrayReference videos to condition the generation on.
6391fe7f61deUpdated: 11/7/2026127.6K runs