Beta interactive avatar model for streaming video with voice-driven actions and live reference-image changes. Full-pipeline and component access share the vidu-s2 selector.
Output resolution720p real-time generation reported in the S2 paper. No general resolution selector in the reviewed streaming request.See specification notes below
Clip durationFull-pipeline real-time session: at most 7,200 seconds. Other routes have separate limits.See specification notes below
Input billingSee ratesVidu API · international · realtime
Output billingSee ratesPublished units and conditions below
Specifications
Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.
Official API announcement date. The current selector description still says beta.
Parameters
Not collected
Architecture
Causal streaming audiovisual generation with Self-Replay Forcing and a one-step latent super-resolution refiner. TurboDiffusion/TurboServe accelerate serving.Source for Architecture (opens in a new tab)
Research-level description. Exact deployed checkpoint size and revision are not disclosed in these sources.
Full-pipeline provider: qwen_omni by default, or doubao. Provider cannot change after creation. Branch voice overrides avatar.voice, then provider default. Component access uses your own ASR/LLM/TTS.Source for Integrated speech providers (opens in a new tab)
These are upstream speech providers, not additional ShengShu-created model checkpoints. Exact upstream model versions are unverified.
Voice controls
Realtime default voice Tina. Language field in voice_model is reserved and ignored. qwen_omni permits session voice updates; doubao does not yet.Source for Voice controls (opens in a new tab)
Voice cloning
Realtime clone route: WAV 16-bit, MP3 or M4A; ≤60 seconds and <10 MB, recommended 10–20 seconds. URL/data-URI field length ≤2,048; voice name 1–16 letters/digits/underscores. Clones are bound to their speech provider; overseas doubao cloning is unsupported.Source for Voice cloning (opens in a new tab)
This is /live/v1/voices/clone, not the different /ent/v2/audio-clone operation. Cloning is not model fine-tuning.
Published international availability is not automatically extended to mainland accounts.
Output link expiry
Optional recording is disabled by default. After session end, an MP4 is usually available within two minutes; download URL is valid 24 hours. This does not establish deletion of input or provider copies.Source for Output link expiry (opens in a new tab)
Live session controls
Full-pipeline idle timeout: 10–7,200 seconds, default 7,200. Optional input/output transcription events. VAD: server or semantic; server silence duration 200–6,000 ms, default 400. Proactive timeout 0 or 500–30,000 ms.Source for Live session controls (opens in a new tab)
Memory and knowledge retrieval
Optional Vidu or external memory/knowledge retrieval. External requests use a configurable endpoint, authorization and timeout up to 30 seconds. Vidu knowledge storage: up to 100 bases/account and 1,000 documents/base. These are service integrations, not a model tool-calling benchmark.Source for Memory and knowledge retrieval (opens in a new tab)
Motion controls
Idle motion supports default/custom/mixed libraries, up to 50 custom entries; execution text 1–200 characters. Interval 5–60 seconds, rest 1–60 seconds. Action libraries: at most 20/account and 100 actions/library.Source for Motion controls (opens in a new tab)
Usage metering
Full-pipeline: three credits per two seconds, rounded up to a two-second block. Component: one credit per second. Offline billing is separate.Source for Usage metering (opens in a new tab)
Published price
International full-pipeline: USD 0.015 per two-second block; component: USD 0.005/second. Mainland: CNY 0.09375 per two-second block and CNY 0.03125/second.Source for Published price (opens in a new tab)
Access & pricing
Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.
General service default: five concurrent tasks per organization; realtime may reject excess sessions with quota errors instead of queueing. Account tier can change the limit.Source for Concurrent requests (opens in a new tab)
Full-pipeline provider: qwen_omni by default, or doubao. Provider cannot change after creation. Branch voice overrides avatar.voice, then provider default. Component access uses your own ASR/LLM/TTS.Source for Integrated speech providers (opens in a new tab)
These are upstream speech providers, not additional ShengShu-created model checkpoints. Exact upstream model versions are unverified.
Output link expiry
Optional recording is disabled by default. After session end, an MP4 is usually available within two minutes; download URL is valid 24 hours. This does not establish deletion of input or provider copies.Source for Output link expiry (opens in a new tab)
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Video generation
Avatar full pipeline1
$0.015
two-second block of live session
Voice generation add-on
Voice cloning add-on2
$4.495
additional voice clone
Service add-on
Concurrency add-on3
$2,080.00
additional concurrent slot per month
Rate conditions
3 credits × USD 0.005 per credit at standard published value. No package discount, promotion or tax included. Starts at on_live when conn_init_ack succeeds. Round session duration upward to a multiple of two seconds; 11 seconds bills 12. Stops at session end.
899 credits × USD 0.005 per credit at standard published value. No package discount, promotion or tax included. First ten realtime clones free per customer. Applies to /live/v1/voices/clone, not standalone /ent/v2/audio-clone. Overseas doubao cloning is unavailable.
416000 credits × USD 0.005 per credit at standard published value. No package discount, promotion or tax included. Five concurrent sessions included per account in the S2 price schedule. Additional resource reservation is separate from usage.
General service default: five concurrent tasks per organization; realtime may reject excess sessions with quota errors instead of queueing. Account tier can change the limit.Source for Concurrent requests (opens in a new tab)
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Video generation
Avatar component1
$0.005
second of live session
Service add-on
Concurrency add-on2
$2,080.00
additional concurrent slot per month
Rate conditions
1 credits × USD 0.005 per credit at standard published value. No package discount, promotion or tax included. Client supplies its own RTC and speech stack; their costs are separate. One credit per billed live second; exact fractional-second settlement is not established.
416000 credits × USD 0.005 per credit at standard published value. No package discount, promotion or tax included. Five concurrent sessions included per account in the S2 price schedule. Additional resource reservation is separate from usage.
General service default: five concurrent tasks per organization; realtime may reject excess sessions with quota errors instead of queueing. Account tier can change the limit.Source for Concurrent requests (opens in a new tab)
Full-pipeline provider: qwen_omni by default, or doubao. Provider cannot change after creation. Branch voice overrides avatar.voice, then provider default. Component access uses your own ASR/LLM/TTS.Source for Integrated speech providers (opens in a new tab)
These are upstream speech providers, not additional ShengShu-created model checkpoints. Exact upstream model versions are unverified.
Output link expiry
Optional recording is disabled by default. After session end, an MP4 is usually available within two minutes; download URL is valid 24 hours. This does not establish deletion of input or provider copies.Source for Output link expiry (opens in a new tab)
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Video generation
Avatar full pipeline1
CN¥0.09375
two-second block of live session
Voice generation add-on
Voice cloning add-on2
CN¥28.09375
additional voice clone
Service add-on
Concurrency add-on3
CN¥13,000.00
additional concurrent slot per month
Rate conditions
3 credits × CNY 0.03125 per credit at standard published value. No package discount, promotion or tax included. Starts at on_live when conn_init_ack succeeds. Round session duration upward to a multiple of two seconds; 11 seconds bills 12. Stops at session end.
899 credits × CNY 0.03125 per credit at standard published value. No package discount, promotion or tax included. First ten realtime clones free per customer. Applies to /live/v1/voices/clone, not standalone /ent/v2/audio-clone. Overseas doubao cloning is unavailable.
416000 credits × CNY 0.03125 per credit at standard published value. No package discount, promotion or tax included. Five concurrent sessions included per account in the S2 price schedule. Additional resource reservation is separate from usage.
General service default: five concurrent tasks per organization; realtime may reject excess sessions with quota errors instead of queueing. Account tier can change the limit.Source for Concurrent requests (opens in a new tab)
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Video generation
Avatar component1
CN¥0.03125
second of live session
Service add-on
Concurrency add-on2
CN¥13,000.00
additional concurrent slot per month
Rate conditions
1 credits × CNY 0.03125 per credit at standard published value. No package discount, promotion or tax included. Client supplies its own RTC and speech stack; their costs are separate. One credit per billed live second; exact fractional-second settlement is not established.
416000 credits × CNY 0.03125 per credit at standard published value. No package discount, promotion or tax included. Five concurrent sessions included per account in the S2 price schedule. Additional resource reservation is separate from usage.
Account-specific training use and input-retention terms remain unverified.
Caching
Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.
01 / INITIAL INPUTProcess & retain
The system processes eligible input and retains reusable state.
02 / LATER REQUESTCheck available state
Reuse depends on matching input, session state and the cache policy.
03 / CACHE HITReuse & generate
Reusing state can reduce repeated work. Any charges follow the billing rules below.
Vidu API · international · realtime
vidu-s2
Not verified
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How it works
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Minimum reusable input
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Lifetime & refresh
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How to enable it
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Usage fields
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Billing rules
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Limits & data boundaries
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026
Vidu API · international · component
vidu-s2
Not verified
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How it works
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Minimum reusable input
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Lifetime & refresh
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How to enable it
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Usage fields
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Billing rules
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Limits & data boundaries
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026
Vidu API · mainland China · realtime
vidu-s2
Not verified
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How it works
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Minimum reusable input
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Lifetime & refresh
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How to enable it
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Usage fields
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Billing rules
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Limits & data boundaries
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026
Vidu API · mainland China · component
vidu-s2
Not verified
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How it works
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Minimum reusable input
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Lifetime & refresh
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
How to enable it
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Usage fields
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Billing rules
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
Limits & data boundaries
Not collected
No public reusable prompt-cache contract or cached-input discount established. Model-internal latent state and retained dialogue memory are different concepts.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
Caching reviewed 14 Oct 2026
Voice choices
122 voices
Choose a voice within Vidu S2 Avatar. These are API voice identifiers, not additional model releases. Source links document each voice.
Published voice catalog
Voice / API ID
Language & accent
Delivery
甜甜 TinaTina
Chinese
Realtime (qwen_omni). A milk-tea girl — sweet and warm International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
林欣宜 CindyCindy
Chinese (Taiwanese accent)
Realtime (qwen_omni). A girl with a soft, coquettish Taiwanese way of speaking International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
清欢 Liora MiraLiora Mira
Chinese
Realtime (qwen_omni). Weaves the warmth of everyday life with her voice International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
知芝 SunnybobiSunnybobi
Chinese
Realtime (qwen_omni). An easy-going, socially awkward girl next door International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
林川野 RaymondRaymond
Chinese
Realtime (qwen_omni). A clear-voiced homebody who lives on takeout International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
晨煦 EthanEthan
Chinese
Realtime (qwen_omni). Standard Mandarin with a slight northern accent. Sunny, warm, energetic, youthful International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
予安 Theo CalmTheo Calm
Chinese
Realtime (qwen_omni). Conveys understanding in silence and heals in speech International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
苏瑶 SerenaSerena
Chinese
Realtime (qwen_omni). A gentle young woman International published catalog; route/provider restrictions apply.Voice source (opens in a new tab)
What the measurements say
No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.
Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
The record behind the model
Why it is included
Qualified closed-weight classification inference from the named Vidu hosted model, official API/research descriptions and operator proprietary-service terms. Public client code and inference libraries are not model weights. Exact private/public checkpoint lineage remains unverified.
