Back to Models

Vidu 2 Reference

zhipuzhipu/vidu2-reference

Zhipu AI Vidu 2 Reference β€” 4s 1280Γ—720 video conditioned on 1+ reference images. Pricing/reference-count details pending ops re-verification against docs.bigmodel.cn.

Vidu 2 Reference is Zhipu AI's reference-image video generation mode in the Vidu 2 family. It accepts one or more reference images plus a text prompt, then creates a short video that preserves subject identity and visual style across the generated motion. The official guide positions Vidu 2 as the speed-and-cost balanced generation, with 4-second 720p output and stronger consistency for multi-reference inputs.

Use the TheRouter model id zhipu/vidu2-reference when you need a lower-cost 720p reference-video route rather than Vidu Q1's higher-resolution generation. The upstream Zhipu API names the model vidu2-reference; TheRouter exposes the same reference-video capability through its routed video-generation surface, with auth, pricing, and supported parameters taken from the live model catalog.

Best for
  • β€’ Character, object, or product clips where reference images must anchor the generated subject identity
  • β€’ High-volume advertising and e-commerce variants that need 4-second 720p video at lower cost than the Q1 generation
  • β€’ Anime shorts, IP spin-offs, and UGC effects where several static assets should share one visual style
  • β€’ Rapid concept tests before spending on higher-resolution or longer video models
Reach for something else if
  • β€’ Pure text-to-video prompts with no reference assets β€” route to a text-to-video model such as zhipu/viduq1-text
  • β€’ First-to-last-frame interpolation where both endpoints are fixed β€” route to zhipu/vidu2-start-end or zhipu/viduq1-start-end
  • β€’ Deliverables that require 1080p output β€” Vidu 2 is documented for 720p; choose a Q1 route instead
  • β€’ Tasks requiring verified snippets today β€” examples stay unproven until a human approves a funded verification run

How TheRouter serves this differently from the vendor

As the vendor operates it

Zhipu's native MaaS API exposes Vidu 2 Reference at /api/paas/v4/videos/generations using upstream model name vidu2-reference and accepts image_url as an array of reference images.

On TheRouter

TheRouter serves the same reference-video capability under public id zhipu/vidu2-reference through its routed video-generation surface with unified TheRouter auth, model catalog pricing, and supported-parameter metadata.

Context Length
--
Max Output
--
Request Priceper request
$0.432/ request

Modalities

imagetext→video

Capabilities

VisionImage to Video

Media Generation Capabilities

video_generation
sizes
  • 1280x720
durations_seconds
  • 4
max_reference_images
3
defaults
size
1280x720
duration_seconds
4

Pricing Breakdown

TypeRate
Request$0.432 / request

Supported Parameters

promptimage_urldurationsizeaspect_ratiomovement_amplitudewith_audio

Specifications

Generation modeReference images to videodocs.bigmodel.cn β†—verified
Input imagesOne or more reference images; the official example passes three image URLsdocs.bigmodel.cn β†—verified
Duration4 secondsdocs.bigmodel.cn β†—verified
Output resolution720p / 1280Γ—720 catalog defaultdocs.bigmodel.cn β†—verified
Upstream model namevidu2-referencedocs.bigmodel.cn β†—verified
Upstream list priceΒ₯2.5 per request in the Zhipu guide; the Z.AI pricing overview currently lists CogVideoX-3 only in its video tabledocs.bigmodel.cn β†—verified
Generated URL validity1 daydocs.bigmodel.cn β†—verified
Concurrent in-flight tasksV0: 5, V1: 10, V2: 15, V3: 20docs.bigmodel.cn β†—verified

Benchmarks

BenchmarkDistributionScoreSource
Official output specification
The official Vidu 2 page lists Reference mode as 4-second, 720p output and shows reference-video requests using duration 4.
4s / 720pdocs.bigmodel.cn β†—
Multi-reference consistency
Zhipu says Vidu 2 strengthens multi-reference consistency so visual style, lighting, and character settings stay aligned across generated shots. No numeric benchmark table is published.
Qualitatively improveddocs.bigmodel.cn β†—
Cost versus Vidu Q1 generation
The official guide describes Vidu 2's 720p output as reduced to 40% of the Q1 version's cost, with Reference mode listed at Β₯2.5 per request.
40% of Q1 cost for 720p Vidu 2 outputdocs.bigmodel.cn β†—

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/videos   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "zhipu/vidu2-reference",
    "prompt": "A slow tracking shot through a neon-lit city street",
    "duration": 8
  }'

Async reference-video task

Submit reference images, a prompt, duration, aspect ratio, and motion controls to create a 4-second 720p video. Snippets are illustrative until human-approved verification is funded.

cURL
curl https://api.therouter.ai/v1/videos \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zhipu/vidu2-reference",
    "prompt": "A cheerful rabbit character drives a tiny car through a sunny market, consistent character design",
    "image_url": [
      "https://example.com/reference-character.jpg",
      "https://example.com/reference-car.jpg",
      "https://example.com/reference-style.jpg"
    ],
    "duration": 4,
    "size": "1280x720",
    "aspect_ratio": "16:9",
    "movement_amplitude": "auto",
    "with_audio": true
  }'

More from zhipu

Similar models

Cross-provider sibling models

News & changes

2025-06-18

Zhipu documents Vidu 2 Reference for 4-second 720p reference-video generation

The Vidu 2 guide describes image-to-video, start-end-frame, and reference-video modes. Reference mode accepts multiple images plus a prompt, with Vidu 2 positioned as a faster, lower-cost 720p generation tier for entertainment, e-commerce, anime shorts, and advertising production.

re-authored by TheRouterdocs.bigmodel.cn β†—

Frequently asked

How is Vidu 2 Reference different from Vidu 2 Start-End?

Reference mode uses one or more images as identity or style references while generating a new short video. Start-End uses exactly two endpoint frames, where the first image is the opening frame and the second image is the final frame.

re-authored by TheRouterdocs.bigmodel.cn β†—
How long and what resolution are Vidu 2 Reference outputs?

The official Vidu 2 guide lists Reference output as 4 seconds at 720p. The TheRouter catalog default for this route is 1280Γ—720, with duration fixed at 4 seconds.

re-authored by TheRouterdocs.bigmodel.cn β†—
Are the snippets on this page verified against the live API?

No. This scheduled curation loop has no paid API budget, and snippet verification requires an operator-set approval flag. Treat the snippets as integration-shaped examples until a human funds and approves verification.

re-authored by TheRouterdocs.bigmodel.cn β†—
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Generation modedocs.bigmodel.cn β†—2026-09-14verified
Input imagesdocs.bigmodel.cn β†—2026-09-14verified
Durationdocs.bigmodel.cn β†—2026-09-14verified
Output resolutiondocs.bigmodel.cn β†—2026-09-14verified
Upstream model namedocs.bigmodel.cn β†—2026-09-14verified
Upstream list pricedocs.bigmodel.cn β†—2026-09-14verified
Generated URL validitydocs.bigmodel.cn β†—2026-09-14verified
Concurrent in-flight tasksdocs.bigmodel.cn β†—2026-09-14verified
Official output specificationdocs.bigmodel.cn β†—2026-09-14verified
Multi-reference consistencydocs.bigmodel.cn β†—2026-09-14verified
Cost versus Vidu Q1 generationdocs.bigmodel.cn β†—2026-09-14verified
Zhipu documents Vidu 2 Reference for 4-second 720p reference-video generationdocs.bigmodel.cn β†—2026-09-14verified
How is Vidu 2 Reference different from Vidu 2 Start-End?docs.bigmodel.cn β†—2026-09-14to verify
How long and what resolution are Vidu 2 Reference outputs?docs.bigmodel.cn β†—2026-09-14to verify
Are the snippets on this page verified against the live API?docs.bigmodel.cn β†—2026-09-14to verify
Help & contact