Wan 3.0 Guide | Venice API Docs

Wan 3.0 is an all-in-one video model, exposed on Venice as a family of variants for text-, image-, and reference-driven generation. It generates up to 30 seconds in a single pass with native audio (dialogue, sound effects, and background music), accepts prompts up to 20,000 characters, and in its reference-to-video form takes up to 20 reference assets at once: 10 images, 5 videos, 5 audio clips, and one document or webpage. The same reference-to-video model handles subject reference, motion and camera reference, voice reference, document-to-video, video editing, and video extension. There is no task field — how you write the prompt decides the workflow. This guide covers the variants, the two input modes, the Venice request parameters, the Omni Reference workflows with their prompt patterns, multimodal limits, pricing, and complete curl examples.

Variants

Every Wan 3.0 tier ships the same three variants. Pick the tier for speed and resolution, and the variant for the kind of input you have.

Tier	Text-to-video	Image-to-video	Reference-to-video	Resolutions	Notes
Wan 3.0	`wan-3-0-text-to-video`	`wan-3-0-image-to-video`	`wan-3-0-reference-to-video`	480p / 720p / 1080p	Standard tier. The only tier that accepts a document or webpage reference.
Wan 3.0 Prime	`wan-3-0-prime-text-to-video`	`wan-3-0-prime-image-to-video`	`wan-3-0-prime-reference-to-video`	480p / 720p / 1080p	Same model and output quality, significantly faster end-to-end, higher per-second price.
Wan 3.0 Pro	`wan-3-0-pro-text-to-video`	`wan-3-0-pro-image-to-video`	`wan-3-0-pro-reference-to-video`	1080p / 2K / 4K	Highest resolution. No 480p/720p rungs.
Wan 3.0 Prime Pro	`wan-3-0-prime-pro-text-to-video`	`wan-3-0-prime-pro-image-to-video`	`wan-3-0-prime-pro-reference-to-video`	1080p / 2K / 4K	Fastest high-resolution tier.

Dotted aliases (wan-3.0-text-to-video, etc.) are accepted for every ID. All variants are async. Submit via POST /api/v1/video/queue, then poll POST /api/v1/video/retrieve until the response body is video/mp4. See Video Generation for the general queue flow.

Two input modes, one model

One model serves two mutually exclusive input modes. On Venice each mode is its own variant, so you never have to worry about mixing them:

Mode	Venice variant	What you supply	Best for
Omni Reference	`*-reference-to-video`	Any mix of reference images, videos, audio clips, and (standard tier) one document or webpage	Most creative work: consistent characters and products, motion transfer, voice-driven dialogue, editing, extension, document-to-video
First & Last Frames	`*-image-to-video`	A strict first frame (`image_url`) and optionally a strict last frame (`end_image_url`)	Shots that must start (and end) on exactly the frame you provide
Text only	`*-text-to-video`	Prompt only	Open-ended generation, fastest to author

Reference images are donors — the model reproduces the subject, style, or scene but does not have to start on that pixel grid. A first frame is a constraint — the video starts on it. If you need the video to open on your exact image, use image-to-video; if you need the same character across shots, use reference-to-video.

Request parameters

Values below are what the Venice API accepts for the Wan 3.0 family. Requests outside these ranges are rejected with a 400 before reaching inference.

Field	Values	Notes
`prompt`	up to 20,000 characters	Required. Long, structured prompts (shot lists, timestamps, per-shot audio) are the intended use.
`duration`	`2s`, `5s`, `10s`, `15s`, `20s`, `25s`, `30s`	Required. With reference videos, input duration + output duration must be ≤ 30 s (see limits).
`resolution`	`480p`, `720p`, `1080p` — Pro tiers: `1080p`, `2k`, `4k`	Optional, default `1080p`.
`aspect_ratio`	`16:9`, `9:16`, `1:1`, `4:3`, `3:4` — plus `adaptive` on image- and reference-to-video	Required. `adaptive` derives the output ratio from your input media and prompt, and is the natural choice for image- and reference-to-video.
`audio`	`true` / `false`	Default `true`. When `false` the output has no audio track. Price is the same either way.
`image_url`	URL or data URL	Image-to-video only: the strict first frame. Shortest side must be ≥ 240 px.
`end_image_url`	URL or data URL	Image-to-video only, optional: the strict last frame.
`reference_image_urls`	array, ≤ 10	Reference-to-video: subject / scene / style donors. Referenced as `@Image1`, `@Image2`, …
`reference_video_urls`	array, ≤ 5	Reference-to-video: motion, camera, or edit/extend source clips. Referenced as `@Video1`, …
`reference_audio_urls`	array, ≤ 5	Reference-to-video: voice timbre, music, or a dance track. Referenced as `@Audio1`, …
`reference_document_urls`	array, ≤ 1	Standard `wan-3-0-reference-to-video` only: one document file or one public webpage URL.

Not supported on this family:

negative_prompt — Wan 3.0 has no negative prompt. Describe what you want instead; for exclusions, state them in the prompt (“no dialogue”, “no background music”, “one continuous take”).
Prompt rewriting — Venice submits your prompt literally, with no automatic prompt expansion, so what you write is what the model sees.
Smart duration — the model can pick its own length, but Venice does not offer that mode because your job is priced up front from the duration you select.

Referencing inputs in the prompt

Refer to reference assets by type and 1-based position: @Image1, @Video2, @Audio1. Numbering follows the order of the arrays you send — reference_image_urls[0] is @Image1, reference_video_urls[0] is @Video1, and so on. The forms Image 1 / Figure 1 are also understood. Mention a reference as many times as you need; each mention is an instruction about how to use it in that part of the shot.

@Image1 walks through the doorway shown in @Image2, then the camera follows in the style of @Video1.
Use the voice from @Audio1 for her line: "We're late."

Omni Reference workflows

All of the patterns below run on wan-3-0-reference-to-video (or the Prime / Pro / Prime Pro equivalent). Include at least one reference image, reference video, or (standard tier) document; audio on its own is not a valid request and must be paired with an image or video.

Subject reference

Lock the appearance of characters, products, and environments across the whole clip. Wan 3.0 is built for pixel-level consistency: every detail of the reference is meant to survive into the output. Prompt pattern:

@Reference subject + action + (dialogue) [+ scene from @ImageN] [+ camera]

Examples:

A continuous, photorealistic cinematic shot. Inside the apartment entrance shown in @Image2, the child from @Image1 stands centre frame, smiles and waves at the camera, then turns and opens the wooden door. Camera fixed; appearance and clothing unchanged throughout.
[Video Type] Game character showcase. [Duration] 15 seconds. Three characters (@Image1, @Image2, @Image3) appear in turn in close-up, each surrounded by their own elemental particles, then demonstrate one signature skill each with matching sound effects.

Tips:

Name the reference every time the subject reappears, especially in multi-shot prompts. The woman is ambiguous once two people are on screen; @Image1 is not.
Keep the reference set lean. Up to 10 images are allowed, but unrelated references compete for attention. Use a character sheet plus the few supporting assets the shot actually needs.
For multi-image scenes, say which image is the subject and which is the scene (“the child from @Image1 … the entrance shown in @Image2”).

Video temporal reference (motion, camera, effects)

A reference video can donate its motion, camera movement, or effects without donating its content. Prompt patterns:

Refer to the camera movement in @Video1, and change the scene to ...
Refer to @Video1, replacing the [subject] with @Image1. Keep the [action], motion trajectory, and background unchanged.
The character in @Image1 performs the same movements and effects as the character in @Video1.

Examples:

Refer to the camera movement style in @Video1, and change the scene to a cup of coffee on a rooftop café table.
Refer to @Video1, replacing the skateboarding man with the woman and clothes from @Image1. Only replace the person; keep the skateboarding action, motion trajectory, and skatepark background completely unchanged.

Audio reference (voice, music, rhythm)

Reference audio can supply a voice timbre for dialogue, a music track to choreograph to, or a sound design reference. Say explicitly how each clip should be used. Prompt patterns:

Extract the tone characteristics from @Audio1 and have the character say: "..."
Dance to the beat of @Audio1: heavy beats → sharp large movements and freeze poses; soft sections → gentle waves.

Example:

Replace the man in @Image1 with the woman from @Video1, sitting by the window on the phone. [Tone and lines] Use the voice from @Audio1 for the line: "Really? That sounds great! When will you be back?" [Mouth and expression] Lip movement must match the generated voice, with natural blinking and small head nods. A 15-second slow push-in, cinematic lighting.

Document or webpage reference (standard tier only)

wan-3-0-reference-to-video can turn a document or a public webpage into a video: a diary into a narrated short film, a spreadsheet into an animated data story, an encyclopedia article into an explainer.

Supply one reference_document_urls entry. Document files (docx, doc, xlsx, xls, pptx, ppt, pdf, txt, key, pages, numbers, md; ≤ 100 MB, as a URL or data URL) are fetched by Venice and forwarded as a file; any other public http(s) URL is forwarded as a webpage link. A file and a link cannot be combined.
Pair the document with the images the video should look like, and tell the model which cells, pages, or sections to draw on.
Prime, Pro, and Prime Pro do not accept documents; a request with reference_document_urls on those tiers is rejected.

Examples:

Help me turn the content of the webpage about quantum mechanics into a fun, popular-science short video in a paper-cut animation style.
Based on the encyclopedia entry, create an animated video titled "Who is Wang Wei" for first-grade students.
A 22-second, four-segment data video on a pure white background, minimalist business style. Segment 1: the title, then the values of cells B7 and G7 slide in with a light-grey arrow between them and the voiceover "In the first half of 2026, monthly GMV grew from 15.8 to 34.6 million dollars." Segment 2: three trend lines drawn left to right from rows 4–6, in the chart style of @Image2 …

Video edit

Edit a single input clip while preserving everything you do not name. Send the clip as reference_video_urls[0] and address it as @Video1. Prompt pattern:

Edit @Video1: [editing target] + [editing action]. [Everything else] remains unchanged.

Sub-patterns:

Edit type	Example prompt
Add elements	`Edit @Video1: Add a boulder on the beach, half-buried in wet sand. As the waves recede, let the water lap over the base of the rock.`
Modify elements	`Edit @Video1: Replace the skateboarding man with a short-haired woman in an athletic tank top and cargo pants. Keep the cap, knee pads, and elbow pads. The tricks, movement trajectory, and skatepark background must remain completely unchanged.`
Remove elements	`Edit @Video1: Remove the woman's sunglasses, revealing her natural eye makeup and gaze. Her movements and the rest of the frame remain unchanged.`
Lighting	`Edit @Video1: Brighten the overall lighting with soft three-point lighting in a cinematic TV-series style. Remove the hard shadow on the right side of the man's face. Movements and frame content unchanged.`
Motion	`Edit @Video1: Swap the positions of the man in the suit and the ostrich. Both keep the same walking pace and rhythm. The street background remains unchanged.`
Style	`Edit @Video1: Change the visual style to claymation.`
Dialogue	`Edit @Video1: Change the male character's dialogue to: "The deal is done. Now... we disappear." Keep his voice timbre and tone.`
Reference-guided	`Edit @Video1: Have the woman put on the hat from @Image1 and the man put on the hat from @Image2. Replace the man's shirt with the denim shirt from @Image3, collar open, sleeves rolled. Everything else unchanged.`
Narrative reshaping	`Reshape the storyline of @Video1: Have the man pick up the guitar and start playing a piece of music.`

Editing bills the input clip duration plus the output duration (see Pricing), and the two together must fit in 30 s.

Video extend

Continue a clip forward, backward, or in both directions. Describe only the new footage — the content, style, and framing are already established by the input. Character descriptions at the top of the prompt help the model keep identities stable across the join. Prompt patterns:

Extend @Video1 by N seconds. [Character A is ...] [Character B is ...] [What happens next, with dialogue.]
Extend @Video1 backward by N seconds. [What happened before.]
Extend N seconds backward from @Video1 as the middle segment: [...]. Extend M seconds forward: [...].

Examples:

Extend @Video1 by 15 seconds. Character A is a man in a dark grey changshan with a prayer-bead bracelet; Character B wears a dark changshan. B sets down his tea bowl and taps the table twice: "The price is negotiable, but when will the goods arrive at port?" A skims the tea foam, leans in: "By the end of the month at the latest." Republic-era business-drama pacing, subdued teahouse ambience.
Extend two seconds backward from @Video1 as the middle segment: the girl walks slowly toward the lens, pauses, and breathes in the cold air. Extend three seconds forward: she looks into the camera, smiles, and raises a gloved hand to catch a snowflake. Golden-hour backlight, shallow depth of field.

Choose duration for the length of the newly generated footage. The input clip’s length plus duration must be ≤ 30 s, so a 15 s input leaves at most 15 s of extension.

First & Last Frames

wan-3-0-image-to-video (and the Prime / Pro / Prime Pro equivalents) starts the video on your image_url and, if given, ends it on end_image_url. Because the frames already fix subject, setting, and style, the prompt should focus on motion and camera:

Prompt = motion description + camera movement

Continue the ink-wash bamboo-forest scene from the first frame. (0:00–0:03) Medium shot; the swordsman rests his hand on the hilt as leaves fall. (0:03–0:08) Push in beneath the hat brim to reveal sharp eyes. (0:08–0:13) Rapid orbit as two ink-silhouette assassins burst from the forest. (0:13–0:20) Handheld fight, then slow motion to a freeze frame on the sheathed blade.
Shot 1 (00:00–00:08): 35 mm film, sunlit rose garden; she touches the brim of her straw hat, slow push-in. Shot 2 (00:08–00:15): close-up, she removes the hat and lowers her head. Shot 3 (00:15–00:23): rear-seat car interior, rain on the window. Shot 4 (00:23–00:30): extreme close-up, she closes her eyes; focus shifts to the raindrops and the frame fades to black.

Use static shot or fixed camera to hold the camera still, and adverbs (slowly, quickly) to control pace. Both images must have a shortest side of at least 240 px.

Prompt guide

Wan 3.0 will make a complete video from one sentence, but output quality tracks prompt completeness. These formulas work on every Venice variant.

Basic formula

For first attempts and open-ended exploration:

Prompt = Subject + Scene + Motion

Advanced formula

Add detail to every slot and layer on aesthetics, style, and sound:

Prompt = Subject + Scene + Motion + Aesthetic control + Stylization + Audio

Subject: appearance in adjectives and short phrases (“a black-haired Miao girl in embroidered ethnic dress”).
Scene: environment, foreground and background, time of day, weather.
Motion: amplitude, speed, and effect (“violently swaying”, “slowly drifting”, “glass shattering”).
Aesthetic control: light source, lighting environment, shot size, camera angle, lens, camera movement.
Stylization: “cyberpunk”, “line-drawing illustration”, “35 mm film”, “claymation”.
Audio: sound effects, background music, dialogue, narration, and voice qualities (“a deep, warm voice”, “bright and clear, American English”).

Sound formula

Wan 3.0 generates audio natively, and the prompt controls it:

Prompt = Voice + Sound effects + Background music
Voice          = "spoken line" + emotion + intonation + speech rate + timbre + accent
Sound effects  = sound source + action + ambient sound
Background music = BGM + style

A man talks about his insomnia. He says, "Love is not getting but giving." Relaxed tone, moderate pace, bright clear voice, American English.
A piece of glass falls from the table onto a wooden floor with a sharp shatter, in a quiet indoor space.
A rainy night in a narrow corridor with a window at the end; suspense-style background music.

Multi-shot formula

For coherent multi-shot narratives with consistent subjects, settings, and mood:

Prompt = Overall description + Shot number + Timestamp + Shot content

A third-person short play about giving up and regaining hope. Shot 1 [0–3 s]: the protagonist sits alone in a corner of the playground reading a letter, then sighs. Shot 2 [4–6 s]: hard cut, fixed camera, close-up on tear-filled eyes. Shot 3 [7–10 s]: hard cut to a plain classroom; a second character in simple clothes smiles warmly and walks toward the first.

Editing formula

Prompt = Editing target + Editing action

Element editing: “The man puts on a red hat”, “Replace the cat with a dog”, “Remove the cat from the grass”. With image references: “Change the man’s clothing to the red jacket in @Image1 and the pants in @Image2.”
Global parameters (style / lighting / colour / weather): “Change the video to an overcast day”, “Convert the video to the 2D cartoon style shown in @Image1.” Add “Keep everything else the same” when you want a narrow change.
Temporal editing (actions, dialogue): “Have the man pick up the cup and take a sip.” For dialogue: “Maintain the character’s voice timbre and tone, and change the girl’s line to: ‘Hello World.’”

Other controls

If you say nothing about shots, dialogue, or music, the model generates them freely. To take control:

Control	How
Number of shots and timing	`Shot 1 [0–3s]: … Shot 2 [3–6s]: …`
Force a single take	`Generate a single shot` or `One continuous take`
Exact dialogue	`XX says: "…"` — the line is preserved verbatim
No dialogue	`No dialogue throughout the entire video.`
No music	`No background music.`
Still camera	`static shot` / `fixed camera`

Multimodal input limits

Values below are what the Venice API accepts for the reference-to-video variants. Requests outside these ranges are rejected with a 400.

Input	Count	Per item	Aggregate	Formats
Reference images	≤ 10	≤ 20 MB, shortest side ≥ 240 px	—	Common image formats, URL or data URL
Reference videos	≤ 5	1–15 s, ≤ 100 MB	≤ 15 s total	`.mp4`, `.mov`
Reference audio	≤ 5	1–15 s, ≤ 15 MB	≤ 15 s total	`.wav`, `.mp3`
Document / webpage	≤ 1	≤ 100 MB	—	docx, doc, xlsx, xls, pptx, ppt, pdf, txt, key, pages, numbers, md, or a public `http(s)` URL. Standard tier only.
Total reference assets	up to 20			10 images + 5 videos + 5 audio + 1 document

Two duration rules apply whenever reference videos are present:

The reference videos together may not exceed 15 seconds.
Reference video duration plus the requested output duration may not exceed 30 seconds. Venice measures the clips when the job is dispatched, so /video/queue still returns a queue_id; an over-budget job then fails on /video/retrieve with a 400 and the pre-charge is refunded. Check clip lengths client-side before submitting.

Image-to-video accepts one image_url and one optional end_image_url, each with a shortest side of at least 240 px.

Pricing

Wan 3.0 is billed per second of output, scaled by resolution. Tiers differ: Prime costs more per second than the standard tier for the same output, and the Pro tiers price 1080p, 2K, and 4K separately. When a request includes reference videos, the reference footage is billed in addition to the output — a 5-second edit of a 10-second clip is charged as 15 seconds. Pass reference_video_total_duration (the sum of all reference clip durations, in seconds) to /video/quote so the quote matches what /video/queue will charge. Call POST /api/v1/video/quote for the authoritative price of a given request shape before submitting it. Prices may change and should not be cached client-side.

curl -X POST https://api.venice.ai/api/v1/video/quote \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "duration": "10s",
    "resolution": "1080p",
    "aspect_ratio": "adaptive",
    "reference_video_total_duration": 8
  }'

audio: false does not change the price.

Complete examples

Text-to-video (30 seconds, multi-shot with audio)

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-text-to-video",
    "prompt": "A modern suspense-style cinematic long take, the whole frame drenched in a pale-green grade. Shot 1 [0-10s]: an empty hospital corridor under cold fluorescent light; the camera glides forward at walking pace, the hum of the lights and a distant trolley the only sounds. Shot 2 [10-20s]: a door at the end opens by itself with a slow creak; the camera does not stop. Shot 3 [20-30s]: inside, a single chair faces a window; rain begins outside, and a low string drone rises. No dialogue throughout the entire video.",
    "duration": "30s",
    "resolution": "1080p",
    "aspect_ratio": "16:9",
    "audio": true
  }'

Image-to-video (first and last frame)

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-image-to-video",
    "prompt": "The woman slowly removes her straw hat and lowers her head, a soft breeze moving her hair. Slow push-in, then hold. 35 mm film texture, warm tones, shallow depth of field.",
    "image_url": "https://example.com/garden-first-frame.png",
    "end_image_url": "https://example.com/garden-last-frame.png",
    "duration": "10s",
    "resolution": "1080p",
    "aspect_ratio": "adaptive"
  }'

Reference-to-video — consistent character in a referenced scene

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "prompt": "A continuous, photorealistic cinematic shot. Inside the apartment entrance shown in @Image2, the child from @Image1 stands centre frame, smiles and waves at the camera, then turns and opens the wooden door with a mechanical click. Warm light floods in from the street shown in @Image3. Camera fixed for the first half, then follows him through the door. Appearance and clothing unchanged throughout; a low cello melody under distant bells.",
    "reference_image_urls": [
      "https://example.com/child.png",
      "https://example.com/apartment-entrance.png",
      "https://example.com/magic-street.png"
    ],
    "duration": "15s",
    "resolution": "1080p",
    "aspect_ratio": "16:9"
  }'

Reference-to-video — voice donor for dialogue

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "prompt": "The woman from @Image1 sits by a rain-streaked window holding a phone to her ear. Use the voice from @Audio1 for her line: \"Really? That sounds great! When will you be back? I miss you so much.\" Lip movement must match the voice, with natural blinking and small nods. A 10-second slow push-in, cinematic lighting.",
    "reference_image_urls": ["https://example.com/woman.png"],
    "reference_audio_urls": ["https://example.com/voice-sample.mp3"],
    "duration": "10s",
    "resolution": "1080p",
    "aspect_ratio": "9:16"
  }'

Reference-to-video — edit an existing clip

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "prompt": "Edit @Video1: Replace the skateboarding man with the woman and clothes from @Image1. Only replace the person; keep the skateboarding tricks, motion trajectory, and skatepark background completely unchanged.",
    "reference_video_urls": ["https://example.com/skatepark.mp4"],
    "reference_image_urls": ["https://example.com/woman-outfit.png"],
    "duration": "10s",
    "resolution": "1080p",
    "aspect_ratio": "adaptive"
  }'

Quote this request with "reference_video_total_duration": 10 (the length of skatepark.mp4).

Reference-to-video — extend a clip forward

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "prompt": "Extend @Video1 by 15 seconds. Character A is the man in the grey changshan; Character B wears the dark changshan. B sets down his tea bowl and taps the table twice: \"The price is negotiable, but when will the goods arrive at port?\" A skims the tea foam and leans in: \"By the end of the month at the latest.\" Unhurried pacing, subdued teahouse ambience.",
    "reference_video_urls": ["https://example.com/teahouse.mp4"],
    "duration": "15s",
    "resolution": "1080p",
    "aspect_ratio": "adaptive"
  }'

Reference-to-video — document to video (standard tier)

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "prompt": "A 20-second, four-segment data video on a pure white background, minimalist business style, in the chart style of @Image1. Segment 1: the title, then the H1 total from the spreadsheet slides in with a bright \"ding\". Segment 2: three trend lines for rows 4-6, drawn left to right month by month. Segment 3: a bar chart of row 5 with each monthly value labelled. Segment 4: a donut chart of the three market shares. Confident, steady female voiceover reading the key numbers; light, fast background music.",
    "reference_document_urls": ["https://example.com/h1-gmv.xlsx"],
    "reference_image_urls": ["https://example.com/chart-style.png"],
    "duration": "20s",
    "resolution": "1080p",
    "aspect_ratio": "16:9"
  }'

Wan 3.0 Pro — 4K text-to-video

curl -X POST https://api.venice.ai/api/v1/video/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-pro-text-to-video",
    "prompt": "West Lake at dawn in thin mist, Chinese rock-colour painting texture with gold-leaf accents. Golden light pierces the clouds in Tyndall beams over a matte azure lake; lotus leaves in the foreground, stone pagodas faint in the fog. One continuous take, slow aesthetic camera, no cuts. Soft guqin and a low ambient wash.",
    "duration": "15s",
    "resolution": "4k",
    "aspect_ratio": "16:9"
  }'

Polling for completion

After every queue submission, save the returned queue_id and poll /video/retrieve until the response body is video/mp4:

curl -X POST https://api.venice.ai/api/v1/video/retrieve \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3-0-reference-to-video",
    "queue_id": "123e4567-e89b-12d3-a456-426614174000"
  }' \
  -o output.mp4

The response is JSON ({ "status": "PROCESSING", ... }) until the job completes, at which point the body switches to video/mp4 bytes. Long, high-resolution jobs take longer; use average_execution_time from the processing response to pace your polling. See Video Generation for the full pattern.

Troubleshooting

`/video/retrieve` returns 400 `Reference video duration (…) plus output duration (…) exceeds this model's 30s combined limit`

Wan 3.0 caps input + output at 30 seconds, and the check runs at dispatch rather than at submit time. A job whose reference clips plus duration exceed 30 s is accepted by /video/queue, then fails on the first /video/retrieve with a 400 and the pre-charge is refunded. The body spells out the numbers and the largest duration that still fits:

{
  "error": "Reference video duration (14.9s) plus output duration (20s) is 34.9s, which exceeds this model's 30s combined limit. Choose an output duration of 15s or less."
}

Pick a shorter duration or trim the reference clips so the two fit in 30 s, and keep the aggregate reference length within 15 s.

`reference_video_urls must have at most 5 videos` / aggregate `> 15s`

Up to 5 reference clips, 1–15 s each, and no more than 15 s in total. Trim client-side before submitting.

`image_url is not supported for text to video models`

Frame inputs belong to the image-to-video variant. Switch to wan-3-0-image-to-video for a strict first frame, or to wan-3-0-reference-to-video and send the image in reference_image_urls if it should act as a donor instead.

`This model does not support end_image_url`

Only the image-to-video variants take a last frame. Reference-to-video has no first/last-frame slot — the two input modes are mutually exclusive.

`This model does not support reference documents`

Document and webpage references are available on wan-3-0-reference-to-video only. Prime, Pro, and Prime Pro reject reference_document_urls.

`A document file and a webpage link cannot be used together`

The single document slot is either a file or a link. Send one or the other.

Image rejected for size

Every image (first frame, last frame, reference) needs a shortest side of at least 240 px. Upscale small assets before submitting.

`The image file is corrupted or unreadable`

The image bytes could not be decoded. This surfaces on /video/retrieve after the job was queued. Check that the file opens locally, that a data URL’s MIME type matches the actual encoding, and re-encode to a plain PNG or JPEG if in doubt.

Output ignores my exclusions

There is no negative_prompt. Put exclusions in the prompt as instructions: “No dialogue throughout the entire video”, “No background music”, “Keep everything else the same”.

Character drifts between shots

Re-state the reference (@Image1) in every shot where the character appears, describe clothing and distinctive features once at the top of the prompt, and keep the reference set to the assets the shot actually needs.

Quote doesn’t match the queued amount

If the request includes reference videos, pass reference_video_total_duration to /video/quote. Reference footage is billed on top of the generated output.

References

Venice video queue endpoint: POST /api/v1/video/queue
Venice quote endpoint: POST /api/v1/video/quote
Companion guide: Video Generation (queue / polling overview)
Companion guide: Reference to Video (Kling O3 + Grok Imagine R2V)
Companion guide: Seedance 2.0 (the other multimodal reference family on Venice)

⌘I

​Variants

​Two input modes, one model

​Request parameters

​Referencing inputs in the prompt

​Omni Reference workflows

​Subject reference

​Video temporal reference (motion, camera, effects)

​Audio reference (voice, music, rhythm)

​Document or webpage reference (standard tier only)

​Video edit

​Video extend

​First & Last Frames

​Prompt guide

​Basic formula

​Advanced formula

​Sound formula

​Multi-shot formula

​Editing formula

​Other controls

​Multimodal input limits

​Pricing

​Complete examples

​Text-to-video (30 seconds, multi-shot with audio)

​Image-to-video (first and last frame)

​Reference-to-video — consistent character in a referenced scene

​Reference-to-video — voice donor for dialogue

​Reference-to-video — edit an existing clip

​Reference-to-video — extend a clip forward

​Reference-to-video — document to video (standard tier)

​Wan 3.0 Pro — 4K text-to-video

​Polling for completion

​Troubleshooting

​/video/retrieve returns 400 Reference video duration (…) plus output duration (…) exceeds this model's 30s combined limit

​reference_video_urls must have at most 5 videos / aggregate > 15s

​image_url is not supported for text to video models

​This model does not support end_image_url

​This model does not support reference documents

​A document file and a webpage link cannot be used together

​Image rejected for size

​The image file is corrupted or unreadable

​Output ignores my exclusions

​Character drifts between shots

​Quote doesn’t match the queued amount

​References

Variants

Two input modes, one model

Request parameters

Referencing inputs in the prompt

Omni Reference workflows

Subject reference

Video temporal reference (motion, camera, effects)

Audio reference (voice, music, rhythm)

Document or webpage reference (standard tier only)

Video edit

Video extend

First & Last Frames

Prompt guide

Basic formula

Advanced formula

Sound formula

Multi-shot formula

Editing formula

Other controls

Multimodal input limits

Pricing

Complete examples

Text-to-video (30 seconds, multi-shot with audio)

Image-to-video (first and last frame)

Reference-to-video — consistent character in a referenced scene

Reference-to-video — voice donor for dialogue

Reference-to-video — edit an existing clip

Reference-to-video — extend a clip forward

Reference-to-video — document to video (standard tier)

Wan 3.0 Pro — 4K text-to-video

Polling for completion

Troubleshooting

`/video/retrieve` returns 400 `Reference video duration (…) plus output duration (…) exceeds this model's 30s combined limit`

`reference_video_urls must have at most 5 videos` / aggregate `> 15s`

`image_url is not supported for text to video models`

`This model does not support end_image_url`

`This model does not support reference documents`

`A document file and a webpage link cannot be used together`

Image rejected for size

`The image file is corrupted or unreadable`

Output ignores my exclusions

Character drifts between shots

Quote doesn’t match the queued amount

References