P-Video-2-Pro

P-Video-2-Pro, based on MiniMax H3, is Pruna’s advanced video generation model, the quality-tier counterpart to P-Video-2. Generate from text, a first-frame image, and an optional last-frame image in one endpoint.

It offers text-to-video and first / last-frame conditioning at 480p and 768p, 24 fps, with duration 5–15 seconds (default 5). Output includes generated audio; audio input is not exposed. Use mode: speed (default) for the faster recipe, or mode: quality for the slower one. prompt_upsampler is independent: off, turbo (default), or max.

Note

When using P-Video-2-Pro , make sure to respect the copyright of the images you use as input and the content you generate.

Pricing: billed per second of the video returned (measured from the finished clip in whole seconds).

Resolution	`mode: speed`	`mode: quality`
480p	$0.02 per second of output video	$0.04 per second of output video
768p	$0.035 per second of output video	$0.075 per second of output video

Speed: model execution time per second of output video.

Resolution	`mode: speed`	`mode: quality`
768p	~0.85 s per second of output video	~1.82 s per second of output video

Tip

Test it now in the P-Video-2-Pro Playground.

Prompt formula

Fast pass

One prompt: subject, action, scene. Enough for first looks. Iterate in speed mode.

[prompt]

subject, action, scene

Locked-in

Add camera, lighting, style, and audio. Optional first-frame or last-frame still for repeatable runs.

[prompt] [image]

subject, action, scene + camera, lighting, style, audio, optional stills for I2V

Fast pass: text-to-video; subject, action, and scene only.

Locked-in: text-to-video; camera, lighting, style, and audio spelled out.

prompt: required. Name the subject, action, and scene. Add camera, lighting, style, and audio when you need a repeatable final.
- Fast pass: “A matte stainless pour-over kettle sits on a seamless light-gray studio sweep. Thin steam rises from the spout.”
- Locked-in: “A narrative film scene, 1970s Roman trattoria at night, warm tungsten, cigarette haze, Super-8 grain. A man and a woman in period clothes sit at a small table with wine. She leans in and quietly says, ‘Then we leave before sunrise.’ The camera slowly dollies around the table. Audio: Italian radio pop from a small speaker, plates, low room tone, her line clear and close.”
image: optional still for image-to-video (jpg, jpeg, png, webp). Keep motion consistent with the reference frame. aspect_ratio is ignored when a reference image is set. Optional last_frame_image for end-frame control.
- Fast pass: omit for text-to-video.
- Locked-in: a high-quality, well-lit still; generate it with P-Image-Ideogram when you need a locked composition.
last_frame_image: optional end-frame still. Use with image when the clip must start and end on specific frames.
- Fast pass: omit.
- Locked-in: pair with image so opening and closing stills stay locked.

Slot	Fast pass (enough to run)	Locked-in (stronger control)
`prompt`	Subject, action, scene.	Subject + action + scene + camera, lighting, style, audio; same wording every rerun.
`image`	Empty (text-to-video).	Well-lit still; canvas follows the reference image.
`last_frame_image`	Empty.	End-frame still when the close must match a packshot or hold.

Tip

For comprehensive video prompting (motion, framing, atmosphere), see the Video Generation guide.

Choosing the right video model

Pruna ships performance video models that share the same prediction API, but each solves a different production problem. P-Video-2-Pro ** generates **new footage (Model: p-video-2-pro).

	P-Video-2-Pro	P-Video-2	P-Video	P-Video-Avatar	P-Video-Animate	P-Video-Replace	P-Video-Edit
One-line job	Generate cinematic footage from prompts	Generate premium footage from prompts	Generate fast / affordable footage	Speak from one still (script or audio)	Retarget one still with clip motion	Swap characters in existing footage	Rewrite content in existing footage
You start with	Text prompt (+ optional first / last frame)	Text prompt (+ optional image / audio)	Text prompt (+ optional image / audio)	Portrait still + `voice_script` or `audio`	Source video + one still	Source video + identity stills	Source video + text prompt
You keep from the source	N/A (new scene); canvas follows reference images when set	N/A (new scene); strong I2V consistency when an image is set	N/A (new scene)	Aspect ratio of the still	Motion, timing, camera from the driver	Camera, timing, blocking, background	Camera, timing, subject performance
Typical ask	“Make an 8 s cinematic product or documentary clip with generated audio.”	“Make a 10 s product ad with music and lip-sync.”	“Draft a 5 s social clip cheaply and fast.”	“This spokesperson says this line in French.”	“Animate this catalog still using our winning ad take.”	“Put our creator in this UGC b-roll.”	“Turn this silver SUV red and keep the camera move.”

Quick decision guide

Need cinematic generation, first + last frame, or 480p / 768p with generated audio → P-Video-2-Pro.
Need 1080p, imported audio, or draft previews → P-Video-2.
Need the fastest / cheapest generation iteration → P-Video.
Single-speaker talking head, voice only, no music bed → P-Video-Avatar.
Footage exists and the hero still should move like the driver → P-Video-Animate.
Footage exists and you need different people in the same shot → P-Video-Replace.
Footage exists and you need a color, product, environment, object, or text edit → P-Video-Edit.

Key features

Marketing and ads: Text-to-video social ads, image-to-video from brand stills, and clips with music or SFX written into the prompt. Prefer P-Video-Avatar only for voice-only talking heads.
Media and entertainment: Music visuals, short-form narrative, and two-shot dialogue with lip-sync. Write the score, ambience, and spoken lines in the prompt; there is no audio upload.
Retail and e-commerce: Product loops and shoppable lifestyle clips. Keep motion subtle and the product readable.
Corporate and education: Training, explainers, and comms with two speakers, SFX, or silent B-roll. Voice-only single-presenter scripts → P-Video-Avatar.
Gaming: Trailers, world reveals, and character moments in a game-engine look. One hero action per clip.
Sports: Documentary and broadcast-style coverage: sideline training, rallies, pack shots, weather and physics. Not a fashion ad.
Generated audio: Output includes generated audio. Write sound or dialogue in the prompt. Audio input is not exposed; for a muxed track, use P-Video-2.
Image-to-video + last frame: Animate a first-frame still; optional last_frame_image for end-frame control. Reference images determine the canvas.
Speed vs quality: mode: speed is the default faster recipe. mode: quality is slower. prompt_upsampler (off / turbo / max) is independent of mode.

Practical constraints

Duration is 5–15 seconds (default 5).
Audio input is not exposed. Write sound or dialogue in the prompt.
aspect_ratio is ignored when a reference image or last_frame_image is provided.
Input images: jpg, jpeg, png, webp.
Resolution is 480p or 768p. Output is 24 fps.
There is no draft mode. mode: speed is the fast recipe; mode: quality is slower.
This model generates new footage. Use P-Video-Edit to modify existing clips.
Rate limit: 250 requests per minute.

Examples

Output

Headphone case social ad · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

A matte-black wireless earphone case sits on wet concrete at night. The lid opens by itself, the headphones lift and unfold, then settle over an unseen wearer’s ears as neon signage reflects in the puddle. Slow orbital camera, premium product lighting, no logos invented, no captions, no voiceover. Audio: low analog synth bed and a soft mechanical click.

A matte-black wireless earphone case sits on wet concrete at night. The lid opens by itself, the headphones lift and unfold, then settle over an unseen wearer’s ears as neon signage reflects in the puddle. Slow orbital camera, premium product lighting, no logos invented, no captions, no voiceover. Audio: low analog synth bed and a soft mechanical click.

Output

1920s Art Deco perfume · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

A 1920s Art Deco vanity at night. No people. A faceted glass perfume bottle, black lacquer, a single ostrich feather, gold geometric inlay. The camera starts locked on the bottle. The stopper lifts by itself and the camera slowly orbits as light breaks through the glass. No logos invented, no captions, no voiceover. Audio: a small 1920s jazz combo, glass clink, no singing.

A 1920s Art Deco vanity at night. No people. A faceted glass perfume bottle, black lacquer, a single ostrich feather, gold geometric inlay. The camera starts locked on the bottle. The stopper lifts by itself and the camera slowly orbits as light breaks through the glass. No logos invented, no captions, no voiceover. Audio: a small 1920s jazz combo, glass clink, no singing.

Output

1990s sneaker hero · text-to-video · 9:16 · 768p · 8s · mode: speed

Example prompt

A 1990s street-style vertical ad. A young man sits on a sunlit stoop, ties a pair of plain white sneakers with no readable logos, then stands and takes two steps toward camera. Handheld but stable, slight film grain, summer haze. He does not speak. No captions. Audio: warm house instrumental, distant traffic, laces pulling, no voiceover.

A 1990s street-style vertical ad. A young man sits on a sunlit stoop, ties a pair of plain white sneakers with no readable logos, then stands and takes two steps toward camera. Handheld but stable, slight film grain, summer haze. He does not speak. No captions. Audio: warm house instrumental, distant traffic, laces pulling, no voiceover.

Output

Acoustic studio session · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

A live acoustic recording session in a small music studio. A woman with an acoustic guitar sings a wordless melody while a violinist plays beside her. Close on fingers on strings, then a two-shot under practical lamps. Photoreal concert coverage for film or TV, not a video game, not an office. Natural faces, subtle lip-sync, no captions, no logos. Audio: intimate folk performance, guitar, violin, small-room reverb, no narrator.

A live acoustic recording session in a small music studio. A woman with an acoustic guitar sings a wordless melody while a violinist plays beside her. Close on fingers on strings, then a two-shot under practical lamps. Photoreal concert coverage for film or TV, not a video game, not an office. Natural faces, subtle lip-sync, no captions, no logos. Audio: intimate folk performance, guitar, violin, small-room reverb, no narrator.

Output

1970s trattoria two-shot · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

A narrative film scene, 1970s Roman trattoria at night, warm tungsten, cigarette haze, Super-8 grain. A man and a woman in period clothes sit at a small table with wine. She leans in and quietly says, "Then we leave before sunrise." He almost smiles, then looks down at his glass. The camera slowly dollies around the table. Realistic faces, subtle lip-sync, natural blinking, no captions. This is live-action cinema, not a video game. Audio: Italian radio pop from a small speaker, plates, low room tone, her line clear and close.

A narrative film scene, 1970s Roman trattoria at night, warm tungsten, cigarette haze, Super-8 grain. A man and a woman in period clothes sit at a small table with wine. She leans in and quietly says, "Then we leave before sunrise." He almost smiles, then looks down at his glass. The camera slowly dollies around the table. Realistic faces, subtle lip-sync, natural blinking, no captions. This is live-action cinema, not a video game. Audio: Italian radio pop from a small speaker, plates, low room tone, her line clear and close.

Output

Film-set rehearsal · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

Behind-the-scenes on a small independent film set in an apartment. Two actors rehearse a quiet scene by a window while a director watches a monitor and a boom operator holds a mic just out of frame. Clapperboard on a table. Photoreal entertainment production, not a video game world, not an office. Natural daylight plus a practical lamp. No captions, no network logos. Audio: whispered rehearsal line, boom rustle, quiet crew, no score.

Behind-the-scenes on a small independent film set in an apartment. Two actors rehearse a quiet scene by a window while a director watches a monitor and a boom operator holds a mic just out of frame. Clapperboard on a table. Photoreal entertainment production, not a video game world, not an office. Natural daylight plus a practical lamp. No captions, no network logos. Audio: whispered rehearsal line, boom rustle, quiet crew, no score.

Output

Studio kettle packshot · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

A photoreal e-commerce packshot. No people. A matte stainless pour-over kettle sits on a seamless light-gray studio sweep. Softbox key from camera left, gentle fill, clean reflections on the metal. The camera starts locked three-quarter, then slowly orbits about 40 degrees as a thin ribbon of steam rises from the spout. Keep the silhouette sharp and the form readable. No logos invented, no on-screen text, no captions. Audio: quiet studio room tone and a faint metal tick, no music, no voiceover.

A photoreal e-commerce packshot. No people. A matte stainless pour-over kettle sits on a seamless light-gray studio sweep. Softbox key from camera left, gentle fill, clean reflections on the metal. The camera starts locked three-quarter, then slowly orbits about 40 degrees as a thin ribbon of steam rises from the spout. Keep the silhouette sharp and the form readable. No logos invented, no on-screen text, no captions. Audio: quiet studio room tone and a faint metal tick, no music, no voiceover.

Output

Apartment coat try-on · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

A photoreal fashion lookbook in a real sunlit studio apartment. A young woman buttons an unstructured camel wool overcoat, turns once in a full-length mirror, then walks two steps toward a window. Handheld but stable, late-afternoon daylight, natural fabric drape and wool texture. No readable brand labels, no captions, she does not speak. Audio: soft lo-fi instrumental, fabric rustle, faint city below the window, no voiceover.

A photoreal fashion lookbook in a real sunlit studio apartment. A young woman buttons an unstructured camel wool overcoat, turns once in a full-length mirror, then walks two steps toward a window. Handheld but stable, late-afternoon daylight, natural fabric drape and wool texture. No readable brand labels, no captions, she does not speak. Audio: soft lo-fi instrumental, fabric rustle, faint city below the window, no voiceover.

Output

Serum UGC demo · text-to-video · 9:16 · 768p · 8s · mode: speed

Example prompt

A photoreal vertical UGC product demo on a real bathroom counter. No face. A pair of hands unscrews a small frosted glass serum bottle, dispenses one drop onto fingertips, and rubs it in. Phone-vertical framing, white tiles, daylight from a side window, condensation on the glass. Natural skin texture, real bottle physics. No logos invented, no captions, no on-screen text. Audio: quiet pop instrumental from another room, bottle click, a distant tap drip, no voiceover.

A photoreal vertical UGC product demo on a real bathroom counter. No face. A pair of hands unscrews a small frosted glass serum bottle, dispenses one drop onto fingertips, and rubs it in. Phone-vertical framing, white tiles, daylight from a side window, condensation on the glass. Natural skin texture, real bottle physics. No logos invented, no captions, no on-screen text. Audio: quiet pop instrumental from another room, bottle click, a distant tap drip, no voiceover.

Output

AAA battlefield trailer · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

AAA video-game cinematic, Unreal Engine 5 real-time look, not live-action film. A playable soldier in grey tactical armor walks forward through volumetric fog on a ruined battlefield. A massive multi-legged insectoid creature looms behind him. Third-person game camera at chest height, shallow depth of field, wet mud. No HUD, no text, no captions, nobody speaking. Audio: low game-trailer score, distant creature roar, boots in mud, no voiceover.

AAA video-game cinematic, Unreal Engine 5 real-time look, not live-action film. A playable soldier in grey tactical armor walks forward through volumetric fog on a ruined battlefield. A massive multi-legged insectoid creature looms behind him. Third-person game camera at chest height, shallow depth of field, wet mud. No HUD, no text, no captions, nobody speaking. Audio: low game-trailer score, distant creature roar, boots in mud, no voiceover.

Output

Cozy farming sim · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

Stylized cozy farming-sim video game, painted 3D, not photoreal live-action. A small farmer in a knit hat waters tomato plants beside a cottage as chickens scatter. Soft rounded textures, gentle bloom, late-afternoon game lighting. No HUD, no text, no captions. Camera slowly orbits at character height. Audio: light folk game soundtrack, chickens, watering can, no voiceover.

Stylized cozy farming-sim video game, painted 3D, not photoreal live-action. A small farmer in a knit hat waters tomato plants beside a cottage as chickens scatter. Soft rounded textures, gentle bloom, late-afternoon game lighting. No HUD, no text, no captions. Camera slowly orbits at character height. Audio: light folk game soundtrack, chickens, watering can, no voiceover.

Output

Cyberpunk rooftop run · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

AAA video-game cinematic, Unreal Engine 5, rainy neon megacity. A playable runner in a scuffed armored jacket sprints across rooftops, vaults a gap, then slides under a holographic billboard of abstract shapes with no real brand names. Third-person game camera tracking from behind, slight motion blur, no HUD, no captions, nobody speaking. Not live-action photography. Audio: synthwave game-trailer score, rain, boots on metal, no voiceover.

AAA video-game cinematic, Unreal Engine 5, rainy neon megacity. A playable runner in a scuffed armored jacket sprints across rooftops, vaults a gap, then slides under a holographic billboard of abstract shapes with no real brand names. Third-person game camera tracking from behind, slight motion blur, no HUD, no captions, nobody speaking. Not live-action photography. Audio: synthwave game-trailer score, rain, boots on metal, no voiceover.

Output

Youth football training · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

Photoreal sports documentary, not a brand ad. A youth football training session on a muddy community pitch at dusk under floodlights. Players in plain unmarked kits pass the ball in a small-sided drill. A coach in a puffer jacket watches from the touchline. Handheld sideline camera at waist height, rain on the lens, breath in the cold. No slow-motion hero shots, no product placement, no captions, no logos invented. Audio: boots on wet grass, a whistle, distant traffic, no music, no voiceover.

Photoreal sports documentary, not a brand ad. A youth football training session on a muddy community pitch at dusk under floodlights. Players in plain unmarked kits pass the ball in a small-sided drill. A coach in a puffer jacket watches from the touchline. Handheld sideline camera at waist height, rain on the lens, breath in the cold. No slow-motion hero shots, no product placement, no captions, no logos invented. Audio: boots on wet grass, a whistle, distant traffic, no music, no voiceover.

Output

Public-court tennis rally · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

Photoreal amateur tennis rally on a worn public hard court at midday. Two players in plain sportswear rally from the baseline. Camera locked behind one baseline, then a small pan to follow the ball. Real sweat, cracked court paint, chain-link fence, no slow-mo, no fashion styling, no captions, no logos invented. Audio: ball on racket, shoes squeak, a distant dog, no music, no voiceover.

Photoreal amateur tennis rally on a worn public hard court at midday. Two players in plain sportswear rally from the baseline. Camera locked behind one baseline, then a small pan to follow the ball. Real sweat, cracked court paint, chain-link fence, no slow-mo, no fashion styling, no captions, no logos invented. Audio: ball on racket, shoes squeak, a distant dog, no music, no voiceover.

Output

Road cycling peloton · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

Photoreal sports documentary of a road cycling peloton on a wet European country road. Tight pack, rain jackets, no readable team names or sponsors. Camera from a motorbike alongside, then lifts slightly as they take a corner. Spray from wheels, tired faces, no roadside ads. No captions. Audio: freehubs, rain, tires on wet asphalt, no music, no voiceover.

Photoreal sports documentary of a road cycling peloton on a wet European country road. Tight pack, rain jackets, no readable team names or sponsors. Camera from a motorbike alongside, then lifts slightly as they take a corner. Spray from wheels, tired faces, no roadside ads. No captions. Audio: freehubs, rain, tires on wet asphalt, no music, no voiceover.

Output

Open-plan office walk-in · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

Photoreal workplace documentary, not a cinema commercial. A contemporary open-plan office on a weekday morning under flat overhead LED panels. Grey carpet, dual monitors with abstract unreadable charts, no brand logos. A woman in a navy knit sweater picks up a laptop from a cluttered desk, walks into a small glass meeting room, and sits opposite a colleague in a pale shirt. They look at the laptop and nod. Handheld but stable, natural skin texture, real office clutter, no slow motion, no cinematic color grade, no captions. Audio: HVAC hum, keyboard clicks, a distant phone, quiet room tone, no music, no voiceover.

Photoreal workplace documentary, not a cinema commercial. A contemporary open-plan office on a weekday morning under flat overhead LED panels. Grey carpet, dual monitors with abstract unreadable charts, no brand logos. A woman in a navy knit sweater picks up a laptop from a cluttered desk, walks into a small glass meeting room, and sits opposite a colleague in a pale shirt. They look at the laptop and nod. Handheld but stable, natural skin texture, real office clutter, no slow motion, no cinematic color grade, no captions. Audio: HVAC hum, keyboard clicks, a distant phone, quiet room tone, no music, no voiceover.

Output

Warehouse safety briefing · text-to-video · 16:9 · 768p · 8s · mode: quality

Example prompt

Photoreal industrial training footage inside a working packaging warehouse. A supervisor in a hi-vis vest shows a pallet jack to a new hire in a plain polo. Forklifts pass in the background. Flat overhead warehouse lighting, scuffed concrete, stacked cardboard. Eye-level follow shot, no drone, no epic lighting. No captions, no logos invented. Audio: warehouse beeps, cardboard, forklift reverse alarm, no music, no voiceover.

Photoreal industrial training footage inside a working packaging warehouse. A supervisor in a hi-vis vest shows a pallet jack to a new hire in a plain polo. Forklifts pass in the background. Flat overhead warehouse lighting, scuffed concrete, stacked cardboard. Eye-level follow shot, no drone, no epic lighting. No captions, no logos invented. Audio: warehouse beeps, cardboard, forklift reverse alarm, no music, no voiceover.

Output

Training room whiteboard · text-to-video · 16:9 · 768p · 8s · mode: speed

Example prompt

Photoreal corporate training room, daytime. A presenter in a grey blazer stands beside a whiteboard with a simple unlabeled flowchart. About ten colleagues sit at tables with notebooks and water bottles. Mixed ceiling lights and window daylight. She points at the board, people take notes. Locked-off tripod shot with a very slow zoom. Realistic faces, no talking-head crop, no captions, no invented company names. Audio: marker on whiteboard, chairs, room tone, no music, no voiceover.

Photoreal corporate training room, daytime. A presenter in a grey blazer stands beside a whiteboard with a simple unlabeled flowchart. About ten colleagues sit at tables with notebooks and water bottles. Mixed ceiling lights and window daylight. She points at the board, people take notes. Locked-off tripod shot with a very slow zoom. Realistic faces, no talking-head crop, no captions, no invented company names. Audio: marker on whiteboard, chairs, room tone, no music, no voiceover.

Integration

P-Video-2-Pro uses the same Pruna predictions API as other performance models. Text-to-video needs only a prompt; image-to-video runs upload files first.

Tip

For more information on how to use the API, see the API Reference.

API endpoint: Base URL: https://api.pruna.ai/v1/predictions

Authentication

-H 'apikey: YOUR_API_KEY'
-H 'Model: p-video-2-pro'

Text-to-video (asynchronous)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Model: p-video-2-pro' \
  -d '{
    "input": {
      "prompt": "A sports car drifting through a neon-lit city at night, cinematic aerial shot",
      "duration": 5,
      "resolution": "768p",
      "aspect_ratio": "16:9"
    }
  }'

Text-to-video (synchronous)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Model: p-video-2-pro' \
  -H 'Try-Sync: true' \
  -d '{
    "input": {
      "prompt": "A sports car drifting through a neon-lit city at night, cinematic aerial shot",
      "duration": 5,
      "resolution": "480p",
      "mode": "speed"
    }
  }'

Image-to-video (first frame)

Upload a reference image, then pass its file URL:

curl -X POST "https://api.pruna.ai/v1/files" \
  -H "apikey: YOUR_API_KEY" \
  -F "content=@/path/to/your/file.jpg"

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H 'apikey: YOUR_API_KEY' \
  -H 'Model: p-video-2-pro' \
  -d '{
    "input": {
      "prompt": "The camera slowly pushes in, the person turns their head and smiles",
      "image": "https://api.pruna.ai/v1/files/fqadqq42xq",
      "duration": 5,
      "resolution": "768p"
    }
  }'

Configuration

Required parameters

Parameter	Type	Description
`prompt`	string	Text prompt for video generation

Optional parameters

Parameter	Type	Default	Description
`prompt_upsampler`	string	`"turbo"`	Expand the prompt before generation: `"off"`, `"turbo"`, or `"max"`. Independent of `mode`
`image`	string (URI)	—	First-frame reference image
`last_frame_image`	string (URI)	—	Last-frame reference image
`duration`	integer	`5`	Video duration in seconds (5–15)
`resolution`	string	`"768p"`	`"480p"` or `"768p"`
`mode`	string	`"speed"`	Generation recipe: `"quality"` or `"speed"`
`aspect_ratio`	string	`"16:9"`	`"16:9"`, `"9:16"`, `"4:3"`, `"3:4"`, `"3:2"`, `"2:3"`, `"1:1"`. Reference images determine the canvas when provided
`seed`	integer	random	Random seed for reproducible generation

Argument recommendations

Use these patterns for consistent quality:

prompt: follow subject / action / scene, then optional camera, lighting, style, and audio. For image-to-video, keep motion consistent with both stills.
image / last_frame_image: high-quality, well-lit stills; aspect_ratio is ignored when a reference image is set.
duration: 5–15 seconds (default 5).
mode: iterate in speed, then rerun finals in quality on the same seed.
prompt_upsampler: leave turbo for production; compare off vs turbo vs max on the same seed before scaling.
resolution: default 768p; use 480p for cheaper, faster drafts.