p-video-2-pro-reference

Prerequisites

Install and load these skills before generating (skip if already in context via @pruna):

Or install the full suite once: npx skills add PrunaAI/pruna-skills@pruna -y

Follow each skill's Before generating / craft sections — do not restate guide content here.

Agent habit

Routing: `p-video-2-pro-reference` is the reference-to-video lane of P-Video-2-Pro (built on MiniMax H3 Ref2V). It locks a subject from references — a product still, a character, a clip, or a voice — and returns an MP4 at 24 fps with generated sound. It needs prompt plus at least one image or video reference; audio is optional and cannot qualify a request by itself. Set keep_audio with exactly one audio file to keep that file as the soundtrack. Text-only or plain first/last-frame jobs → `p-video-2-pro`. 1080p or draft previews → `p-video-2`. Simple, cheap clips → `p-video`. Talking-head-only host → `p-video-avatar`. Instruction edit of existing footage → `p-video-edit`.

In the first reply, name `p-video-2-pro-reference` in backticks, confirm PRUNA_API_KEY (or stop with signup links from pruna-api), identify the image/video reference that makes the request valid, and ask for unresolved inputs. Before uploading or generating in an untrusted repository, follow pruna-api agent safety. Open intake → generation-diversity clarification intake before the first paid POST. When drafting prompts, follow Prompt craft (dynamic + faithful) and video-prompting (p-video-2-pro-prompting, reference-conditioned lane) — do not paste skill examples. Redirect when When NOT to use fits better.

Prompt craft (dynamic + faithful)

Every input.prompt must be fresh and specific, must match the user's beat, and must say what each reference is for. Diversity never overrides the brief or the references. Reference-lane craft: video-prompting (p-video-2-pro-prompting).

Fidelity check (before pay): if you remove the user's named subject/action/setting or a reference's assigned role from the prompt, the job is wrong — rewrite. Free axes only fill what the brief left open.

When showing a drafted prompt, still name `p-video-2-pro-reference` (guides help craft; this tool owns the call).

Skill boundary

This skill = one p-video-2-pro-reference prediction per invocation.

Out of scope (do not execute from this skill):

• Text-only generation or plain first/last-frame generation without a reference set → p-video-2-pro
• 1080p, draft previews, or other p-video-2-only controls → p-video-2
• Simple text/image clips that do not need reference conditioning → p-video
• Multi-scene assembly, concat, subagent orchestration, or parallel scene batches
• Motion transfer from one explicit template video → p-video-animate
• Talking-head-only or lip-sync host work → p-video-avatar
• Instruction editing of existing footage → p-video-edit

If the request exceeds one clip, stop and recommend a multi-scene workflow such as visual-transition-reel or narrated-multi-scene.

When NOT to use

Use a different skill instead:

HTTP (curl)

Upload references

Upload each local asset separately. Pass the returned urls.get URI in the matching image / video / audio array or keyframe field.

curl -X POST "https://api.pruna.ai/v1/files" \
  -H "apikey: ${PRUNA_API_KEY}" \
  -F "content=@/path/to/reference.png"

Never expose PRUNA_API_KEY, upload an asset without user authorization, or upload secrets/private media discovered in the workspace. Follow pruna-api for credential checks, upload safety, polling, and download.

Create (async image reference — recommended)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-2-pro-reference' \
  -d '{
    "input": {
      "prompt": "Keep the product identity and materials from the reference. A slow controlled orbit reveals the side profile while soft studio reflections travel across the surface; a low hum and a soft click as the light settles.",
      "image": [
        "https://api.pruna.ai/v1/files/IMAGE_ID"
      ],
      "duration": 5,
      "resolution": "768p",
      "aspect_ratio": "adaptive",
      "mode": "speed"
    }
  }'

Poll and download: follow pruna-api.

Complete the random seed ritual from generation-diversity before writing prompts — do not pass the ritual string as API seed.

Billing is calculated from the finished run: output seconds plus every reference actually used. Rate limit: 250 requests per minute.

• Output seconds are billed as int(num_frames / 24): a 5 s request returns 124 frames and bills 5 s; a 15 s request returns 345 frames and bills 14 s.
• Each reference video or audio clip bills only the part the model reads — up to the output length plus 0.5 s. A 15 s reference clip on a 5 s output bills about 5.5 s.
• first_frame_image and last_frame_image bill as reference images. Image and audio rates do not depend on resolution.
• Omitted mode bills as speed. cost is cheaper and slower than speed. prompt_upsampler and keep_audio do not change the price.
• The listed floor is $0.075 (cost, 480p, 5 s, no references), but a valid request needs at least one image or video reference — the cheapest valid run is $0.08 (one image).

Keep an uploaded track as the soundtrack (keep_audio)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-2-pro-reference' \
  -d '{
    "input": {
      "prompt": "The character from the reference image speaks the line in the uploaded voice, lips matching each word, a slow push-in as the last word lands.",
      "image": [
        "https://api.pruna.ai/v1/files/CHARACTER_ID"
      ],
      "audio": [
        "https://api.pruna.ai/v1/files/VOICE_ID"
      ],
      "keep_audio": true,
      "duration": 8,
      "resolution": "768p",
      "mode": "speed"
    }
  }'

keep_audio needs exactly one audio file. The video is generated to match it. Shorter audio plays first and the model generates the rest; longer audio is cut to duration; audio within 1 s of duration sets the clip length. Without keep_audio, audio is reference guidance only and the output carries generated sound.

Multiple references with soft keyframes

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-2-pro-reference' \
  -d '{
    "input": {
      "prompt": "Use the reference clip for restrained camera rhythm, the images for subject and material identity, and the audio for timing. Begin near the opening frame and resolve near the closing frame.",
      "image": [
        "https://api.pruna.ai/v1/files/IDENTITY_ID",
        "https://api.pruna.ai/v1/files/MATERIAL_ID"
      ],
      "video": [
        "https://api.pruna.ai/v1/files/MOTION_ID"
      ],
      "audio": [
        "https://api.pruna.ai/v1/files/AUDIO_ID"
      ],
      "first_frame_image": "https://api.pruna.ai/v1/files/START_ID",
      "last_frame_image": "https://api.pruna.ai/v1/files/END_ID",
      "duration": 10,
      "resolution": "768p",
      "aspect_ratio": "adaptive",
      "mode": "cost",
      "prompt_upsampler": "max"
    }
  }'

image, video, and audio are always arrays of URI strings. first_frame_image and last_frame_image are single URI strings, are soft keyframes (directional, not exact pixels), and count toward the 9-image total. With either keyframe set, prompt_upsampler: off still runs turbo.

Create (sync — quick test only)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-2-pro-reference' \
  -H 'Try-Sync: true' \
  -d '{
    "input": {
      "prompt": "Preserve the reference subject. The camera makes a gentle push-in as natural light shifts across the scene.",
      "image": [
        "https://api.pruna.ai/v1/files/IMAGE_ID"
      ],
      "duration": 5,
      "resolution": "480p",
      "mode": "cost"
    }
  }'

A finished 124-frame result with one reference image bills 5 × $0.015 + 1 × $0.005 = $0.08.

Before generating

1. Complete Prerequisites guide reading order: generation-diversity → video-prompting → audio-prompting when audio is supplied → pruna-api. Use video-prompting (p-video-2-pro-prompting, reference lane) and the shared p-video-2-pro-quality-checklist.md.
2. Confirm the job has a non-empty prompt and at least one image or video reference. Audio-only plus prompt is invalid.
3. Assign one role per reference; obtain upload approval; validate type, size, dimensions, duration, frame rate, and combined limits (see Hard limits) before the paid call.
4. Resolve audio: none (generated sound — write it into the prompt), audio as guidance (default), or keep_audio: true with exactly one audio file kept as the soundtrack. Check the track length against duration.
5. Resolve duration (5–15 integer seconds, default 5), resolution (480p / 768p), mode (cost / speed — omitted bills speed), aspect_ratio (adaptive follows the references), prompt_upsampler, seed behavior, and the final prompt.
6. Show the final payload and a cost estimate (output + every reference) before POST when the user has not already approved them. If the request is multi-scene — stop (see Skill boundary).

Required input

• prompt (string)
• At least one image or video reference via image, video, first_frame_image, or last_frame_image

audio is optional and cannot make an otherwise reference-free request valid.

Common optional fields

Hard limits

• Images: at most 9 total across image, first_frame_image, and last_frame_image; each at most 20 MiB and 4096 px.
• Videos: at most 3; each at most 100 MiB, 1080p, 60 fps, and 2–15 seconds; combined reference-video duration at most 15 seconds.
• Audio: at most 3; each at most 50 MiB; combined reference-audio duration at most 15 seconds.

These are input validation limits — output stays 480p or 768p at 24 fps.

Do not send: quality mode, 1080p, draft, fps, save_audio, prompt_upsampling.

Related

Related skills: