Vidu (ShengShu Technology) AI Comic Drama Platform Research

1. Introduction

Vidu was released by Beijing ShengShu Technology Co., Ltd. together with Tsinghua University, and is officially positioned as "China's first long-duration, high-consistency, high-dynamics video foundation model". Among the 8 objects in this group, Vidu is the commercial platform with the most explicit and complete L4 (memory and state) form: its "Subject Bank" saves characters, props, and scenes as persistent entries that can be referenced with one click, and multi-subject consistency supports up to 7 subject images.

This directly hits the group's core thesis — AI comic drama is the scenario with the greatest L4 pressure among the six Harness layers; whoever makes L4 solid can push AI comic drama from "being able to generate" to "being able to produce continuously." Through the "Reference-to-Video" paradigm, Vidu turns the consistency problem from a model-capability problem into an asset-engineering problem, making it the most representative L4 sample in this group.

1.1. Developer and Capital

Financing and capitalization:

Strategic roadmap: foundation world model → world generation model WGM (digital world) + world action model WAM (physical world); in December 2025, the world action model Motus was open-sourced.

1.2. Positioning

Vidu Q3 is positioned as "made for drama" — industrialized content production, full Reference-to-Video capability matrix, native audio-visual output, and commercial delivery ready.

This positioning is the closest in this group to the real needs of AI comic drama: it does not pursue the breadth of general creative video, but concentrates resources on the industrialized delivery of the single content category of "episodic drama." The three-layer capability matrix (VFX, sound effects, and scenes) also confirms this.

1.3. Pricing

Vidu officially claims pricing of "1/3 of the industry average."

Third-party API aggregator pricing (Atlas Cloud, [To be verified: third-party API aggregator pricing]), Vidu Q3 per-second prices:

C-end pricing gap: viduq3.net is a third-party site (one-time credit packs of $9.9/$29.9/$49.9/$99.9); the vidu.cn official site shows no complete price list. This group was unable to obtain Vidu's official C-end pricing page, marked [To be verified].

Other cost-related points: off-peak mode with unlimited credits (unlimited free video generation during off-peak hours).

1.4. Open Forms

2. Glossary

3. Feature Description

3.1. Reference-to-Video

Reference-to-Video is Vidu's core paradigm, officially described as "Everything Referenceable" — extracting characters, scenes, and wardrobe-and-props as reusable modeling material. Upload 3 or more reference images, and the model models them as reusable subjects and keeps them consistent in the output.

Comparison with other objects in the group:

Vidu's characteristics are clear reference count and clear subject concept. The "subject" abstraction is key — it unifies characters, props, and scenes into the same class of referenceable objects, rather than building separate "character libraries" and "scene libraries."

3.2. Subject Bank and Multi-Subject Consistency

Subject Bank: saves characters, props, and scenes in the Subject Bank, with one-click selection of reference subjects.

Multi-subject consistency: upload up to 7 subject images and keep multiple subjects consistent within a single generation.

This is the only asset-library mechanism in this group that explicitly discloses a capacity limit (7 images), and the only one that explicitly covers three asset categories (characters, props, and scenes).

3.3. VFX and Sound-Effect Capability Matrix

Vidu Q3's three-layer capability matrix:

The scene layer explicitly lists "comic drama", making Vidu the only platform in this group that treats comic drama as an independent scene category in its capability matrix.

3.4. Camera Control and Scene-Boundary Detection

• Native camera control: frame-level director commands for precise control of camera movement.
• Smart Cut: automatic scene-boundary detection and transitions.

Together, these two provide the bidirectional capability of "automatically splitting long material into shots" and "precisely controlling camera movement within a shot."

3.5. Other Capabilities and Templates

4. Platform Architecture

图 4-1｜Vidu 平台六层架构：从 U-ViT 底座到主体库与商业化分层

数据来源：基于本文分析绘制的示意图。

4.1. Overall Architecture

┌────────────────────────────────────────────────────────────────┐
│  商业化分层  SaaS：Vidu Agent · Vidu Claw（个人/小工作室）         │
│              MaaS：Vidu AI 开放平台 · Vidu.API（企业/开发者）      │
│              云平台：阿里云百炼                                    │
├────────────────────────────────────────────────────────────────┤
│  编排与控制  原生镜头控制（帧级导演指令）                           │
│              Smart Cut（场景边界检测与转场）                       │
├────────────────────────────────────────────────────────────────┤
│  能力矩阵  特效层 6 · 音效层 5 · 场景层 4（含漫剧）                │
├────────────────────────────────────────────────────────────────┤
│  资产层  主体库（角色/道具/场景）· 多主体一致性（最多 7 张）         │
├────────────────────────────────────────────────────────────────┤
│  模型层  Vidu Q1（叙事逻辑）· Q2（AI 演技）· Q3（工业化生产）       │
│          Vidu S1（实时交互）· Motus（世界行动模型，2025-12 开源）   │
├────────────────────────────────────────────────────────────────┤
│  底座  U-ViT 架构（Diffusion + Transformer 融合）                 │
└────────────────────────────────────────────────────────────────┘

4.2. Model Layer

The Q1→Q2→Q3 evolution path is worth noting: first solve "whether the story can advance" (narrative logic), then "whether characters can perform" (micro-expressions and body language), and finally "whether it can be delivered industrially" (Reference-to-Video capability matrix + native audio-visual output + commercial readiness). This is a clear route from content to engineering.

4.3. Asset Layer

The Subject Bank is the core differentiator of this architecture; see the L4 analysis in Section 5.4.

4.4. Commercialization Layers

SaaS layer: Vidu Agent, Vidu Claw, for individuals / small studios. Vidu Claw is positioned as "hire AI creative staff."

MaaS layer: Vidu AI Open Platform, Vidu.API, for enterprise batch production and developer integration. It is available on the Alibaba Cloud Bailian platform.

This layering lets the same set of model capabilities be used directly by creators and also integrated by enterprises into their own production systems.

5. Harness Design

5.1. L1 Context Engineering Layer

Vidu's L1 mechanism is Reference-to-Video (Everything Referenceable): extracting characters, scenes, and wardrobe-and-props as reusable modeling material and injecting them into the context in the form of reference images.

Compared with other objects in the group, Vidu's L1 has two characteristics:

1. Reference-as-modeling: the official wording emphasizes "modeled as a reusable subject," meaning the reference image is not a simple conditional input but is modeled into a subject representation. This goes a step beyond "pasting an image on."
2. Clear capacity: up to 7 subject images. Although fewer than Seedance (9 images + 3 videos + 3 audio) and Hailuo (12 reference files), the explicitly publicized limit itself is engineering-friendly — users can plan deterministically.

Gaps:

1. No public context-compression mechanism (compared with Hailuo H3-Context-IR's 25:1 compression).
2. Whether references support video and audio modalities is not clearly explained (compared with Seedance and Hailuo, which explicitly support them), marked [To be filled].
3. No public context-priority sorting or cache-reuse mechanism.

5.2. L2 Tools and Execution Layer

Vidu's L2 is organized as a capability matrix, the most structured tool layer in this group:

Assessment: Vidu's L2 ranks "strong." The 5-major sound-effect layer is unique in this group (other platforms mostly do post-production dubbing or native audio without sound-effect categorization). The 6-major VFX layer covers the most common visual needs of comic drama.

Programmability: provides Vidu.API (MaaS) and Alibaba Cloud Bailian integration, with clear programmability. However, no CLI or Skills-type tools were found (compared with PixVerse), marked [To be filled].

5.3. L3 Orchestration and Control Layer

Vidu has three clear threads at L3:

1. Native camera control: frame-level director commands. This is the finest-grained camera control in this group — not a coarse-grained choice of "wide / medium / close-up" but precise frame-level specification.
2. Smart Cut: automatic scene-boundary detection and transitions. This solves the problem of automatically splitting long material into shots.
3. Vidu Agent / Vidu Claw: Agent-like orchestration at the SaaS layer; Vidu Claw is positioned as "hire AI creative staff."

Assessment: Vidu's L3 ranks "strong." Its orchestration approach differs from both Kling (storyboard commands) and Hailuo (timeline) — it is a combination of frame-level precise control + automatic boundary detection, offering both precision and automation.

Gaps: no editable node graph (compared with PixVerse Canvas, ComfyUI); no public conditional-branch and batch-processing mechanisms; whether orchestration results can be exported and versioned is not explained, marked [To be filled].

5.4. L4 Memory and State Layer

Vidu's L4 is the strongest among the commercial platforms in this group, and the best illustration of the group's core thesis.

Why this matters: the cross-shot, cross-episode consistency of AI comic drama is essentially a cross-session state-persistence problem. Vidu's Subject Bank productizes this — users do not need to build their own asset-management system; the platform provides:

• State carrier (subject entries)
• State naming (three categories: characters / props / scenes)
• State reference (one-click selection)
• State capacity (up to 7)

This is the only commercial platform in this group that makes all four L4 elements explicit.

Remaining gaps:

1. Whether the 7-image capacity limit is enough: for multi-character ensemble scenes, 7 subject images (including characters, props, and scenes) may be tight. Compared with Hailuo's 12 reference files and Seedance's 9 images + 3 videos + 3 audio, Vidu's capacity is on the small side.
2. No plot state machine: the Subject Bank solves "looking the same" but not "plot-state continuity." Cross-episode character relationships, injuries, and timeline progression still need external maintenance.
3. Cross-project reuse rules not disclosed: whether the Subject Bank can be reused across projects and teams is not explained, marked [To be filled].

5.5. L5 Evaluation and Observation Layer

Vidu's L5 has the most targeted evidence in this group of commercial platforms.

The SuperClue result is especially critical: it is the first dedicated Reference-to-Video capability benchmark, and "reference-to-video" is precisely the core technical path for AI comic drama consistency. Vidu's double first-place finish on this benchmark shows third-party endorsement of its L4 capability — it is the only platform in this group with a third-party benchmark specifically targeting "consistency."

Gaps: no platform-built-in availability-statistics dashboard, regression set, or trajectory tracking, marked [To be filled].

5.6. L6 Governance and Safety Layer

Vidu's L6 mechanism is ecosystem layering: SaaS (individuals / small studios) and MaaS (enterprise batch production / developer integration).

Assessment: Vidu's L6 ranks "medium." The ecosystem layering itself provides a division of usage boundaries, but explicit permission, quota, and audit mechanisms are lacking (compared with PixVerse Team Plan's RBAC + credit cap + usage analytics + unified billing).

Gaps:

1. No public explanation of AI-generated synthetic-content labeling found — whether it meets the explicit-labeling and AIGC metadata implicit-labeling requirements of GB 45438—2025 has no public result, marked [To be filled].
2. No real-person image verification mechanism found — compared with Doubao/Seedance, Vidu has no public information on this red line.
3. No public team-collaboration and permission-management capability — compared with PixVerse Team Plan.
4. C-end pricing is opaque — no complete price list on the official site.
5. Commercial-licensing rules not disclosed — "commercial delivery ready" is Q3's positioning statement, but the specific commercial-licensing terms have no public explanation.

5.7. Six-Layer Capability Matrix

6. Case Studies

6.1. Case 1: Industrialized Episodic Production

• Background: the core requirement of AI comic drama is "100 episodes feel like one drama" — that is, cross-episode consistency of characters, props, and scenes.
• Approach: use Vidu Q3's Reference-to-Video and Subject Bank — register main characters, key props, and fixed scenes in the Subject Bank, and one-click select reference subjects for each generation; pair with native camera control for frame-level directing and Smart Cut for scene-boundary detection and transitions.
• Results: Vidu Q3 won first place on both the multi-image and single-image reference leaderboards with a decisive advantage in SuperClue, the first dedicated Reference-to-Video capability benchmark, and topped the Artificial Analysis overall ranking at its January 2026 release. Specific episodic-project delivery data has not been publicly disclosed, marked [To be filled].
• Harness interpretation: this is the only solution in this group backed by a "consistency-specific third-party benchmark." For teams whose first priority is continuous production, Vidu's L4 is currently the most effort-saving choice — the platform has already built the asset library.

6.2. Case 2: Enterprise-Grade Customer Collaboration

• Background: top customers in the visual-effects and animation field have the highest requirements for consistency and delivery standards.
• Approach: Vidu's SaaS (Vidu Agent, Vidu Claw) + MaaS (Vidu AI Open Platform, Vidu.API) layering, combined with the already-available Alibaba Cloud Bailian platform.
• Results: customers and partners include Sony Pictures, Tencent Animation, and China Literature (Yuewen); users and business cover over 200 countries and regions worldwide; in 2025 users and revenue grew more than 10x. Specific project details have not been publicly disclosed, marked [To be filled].
• Harness interpretation: with China Literature as an online-novel IP party and Tencent Animation as an animation content party, the collaboration with Vidu points to the "novel/comic IP → comic drama" conversion chain, directly connecting to this group's 05-AI-Novel scenario.

6.3. Case 3: Cost Optimization in Off-Peak Mode

• Background: AI comic drama has large production-demand, but most teams face tight cash flow; the industry payback rate is below 1.3%, so cost sensitivity is extremely high.
• Approach: Vidu provides off-peak mode with unlimited credits — unlimited free video generation during off-peak hours.
• Results: the specific definition of off-peak hours and concurrency limits have not been publicly disclosed, marked [To be filled].
• Harness interpretation: off-peak mode is a typical L6 cost-control measure (peak shaving and valley filling), following the same approach as PixVerse's Off-Peak / Preview modes in this group. For non-urgent batch generation tasks, queueing tasks into the off-peak queue can significantly reduce marginal cost — a scheduling strategy worth adopting from an engineering standpoint.

7. Summary

7.1. Strengths

1. Most explicit and complete L4 form: Subject Bank (characters/props/scenes) + multi-subject consistency (up to 7), and the only commercial platform in this group that makes all four asset-library elements (carrier, naming, reference, capacity) explicit.
2. Consistency backed by a dedicated third-party benchmark: first on both the multi-image and single-image reference leaderboards in SuperClue — the only third-party benchmark result in this group aimed specifically at "consistency."
3. Most precise "made for drama" positioning: the scene-layer capability matrix explicitly lists "comic drama," and the VFX and sound-effect layers are the most structured in this group.
4. Frame-level camera control + Smart Cut: offers both precision and automation.
5. Solid technology foundation: the U-ViT architecture is the world's first architecture fusing Diffusion with Transformer; backed by the Tsinghua Zhu Jun team.
6. Clear ecosystem layering: SaaS (individuals) + MaaS (enterprises) + Alibaba Cloud Bailian, covering users of different scales.
7. Strong capital and customer validation: nearly RMB 2 billion Series B (led by Alibaba Cloud), valuation over USD 2 billion, customers including Sony Pictures, Tencent Animation, and China Literature, with 2025 users and revenue growing more than 10x.
8. Off-peak mode: friendly to cost-sensitive teams.

7.2. Limitations

1. Reference capacity on the small side: up to 7 subject images, fewer than Hailuo's 12 reference files and Seedance's 9 images + 3 videos + 3 audio. Multi-character ensemble scenes may be constrained.
2. No context-compression mechanism: compared with Hailuo H3-Context-IR's 25:1 compression, Vidu follows the "clear capacity + modeling reuse" route, so capacity is a hard limit.
3. No plot state machine: the Subject Bank solves appearance consistency, not plot-state continuity.
4. C-end pricing is opaque: no complete price list on the official site, only third-party-site quotes.
5. Insufficient L6 information: AI labeling, real-person image verification, team collaboration permissions, and commercial-licensing terms all lack public explanation.
6. No CLI/Skills: programmability is weaker than PixVerse.
7. Single-run duration of 16 seconds: on par with most platforms in this group (15~16 seconds); long shots require stitching.
8. Maximum resolution 1080p: the only commercial platform in this group not reaching 2K/4K (compared with Kling 4K, PixVerse 4K, Hailuo 2K, Seedance 4K).

7.3. Applicability Boundary

7.4. Selection Recommendations

• The core reason to choose Vidu is L4, not image quality or resolution. If your primary problem is "characters look different in episode 1 and episode 47," Vidu's Subject Bank is currently the most direct answer.
• Treat the Subject Bank as your main asset library, but not your only one: the 7-subject-image limit means you need to prioritize your assets — main characters and fixed scenes go into the Subject Bank, while secondary props and external references are carried by external systems.
• Be sure to build your own plot state machine: the Subject Bank only handles appearance. Cross-episode character relationships, injuries, and timeline progression need to be maintained externally as structured state and injected into the prompt as text at each generation.
• Use off-peak mode for batch generation: queueing non-urgent batch shots into the off-peak queue can significantly reduce marginal cost. We recommend dividing generation tasks into two queues: "urgent manual refinement" and "off-peak batch."
• Confirm resolution in advance: Vidu maxes out at 1080p; if final delivery requires 2K/4K, you need to do upscaling in post, or mix in other platforms (e.g., Hailuo's In-Context Regeneration 2K).
• Verify compliance capabilities yourself: there is no public explanation of AI-labeling and real-person-image mechanisms; works distributed on domestic platforms must build their own pipeline meeting GB 45438—2025's explicit-labeling and AIGC metadata implicit-labeling requirements.
• Request an official quote before C-end purchasing: the official site has no complete price list, so third-party-site quotes ($9.9/$29.9/$49.9/$99.9 credit packs) cannot be used as a budgeting basis.

Information-Gap Statement

1. Vidu's official C-end pricing: viduq3.net is a third-party site (one-time credit packs of $9.9/$29.9/$49.9/$99.9); the vidu.cn official site shows no complete price list.
2. Whether reference materials support video and audio modalities: officially "up to 7 subject images," whether video/audio references are supported is not explained.
3. Cross-project / cross-team reuse rules for the Subject Bank: no public explanation.
4. Specific rules of off-peak mode: the definition of off-peak hours, concurrency limits, and whether there is a generation-quality difference have not been disclosed.
5. How Vidu implements AI-generated synthetic-content labeling: whether it meets the explicit-labeling and AIGC metadata implicit-labeling requirements of GB 45438—2025 has no public result.
6. Vidu's real-person image verification mechanism: whether real-person verification and real-face restrictions exist has no public result.
7. Vidu's team-collaboration and permission-management capability: whether RBAC, credit caps, and usage analytics similar to PixVerse Team Plan exist has no public result.
8. Specific commercial-licensing terms: Q3 is positioned as "commercial delivery ready," but the licensing scope, restrictions, and responsibility division have no public explanation.
9. Vidu.API's capability coverage: which web-side capabilities can be called through the API has no complete comparison explanation.
10. Technical specifications of native camera control: the specific syntax of "frame-level director commands" and the supported control dimensions (push-pull, par-away-tilt, pan-tilt, depth-of-field, focal length, etc.) have no public explanation.
11. Data source for the 16-second single-run duration and 1080p: from a third party, Atlas Cloud, not confirmed on the Vidu official site.
12. Platform-built-in evaluation capability: whether availability statistics, regression sets, and trajectory tracking exist has no public result.

8. References

1. Vidu official site — https://www.vidu.cn/
2. Tencent Cloud Developer Community "Vidu Q3 Reference-to-Video Review" — https://cloud.tencent.com/developer/article/2655977
3. Atlas Cloud "ShengShu Models" — https://www.atlascloud.ai/providers/shengshu
4. Atlas Cloud "Vidu Q3 AI Video Generator Now on Atlas Cloud" — https://www.atlascloud.ai/it/blog/ai-updates/vidu-q3-ai-video-generator-now-on-atlas-cloud-create-16s-cinematic-with-native-audio-sync
5. AI-Pirates "Vidu: AI Video with Reference Consistency" — https://www.ai-pirates.com/en/glossar/vidu
6. China Business Journal "ShengShu Technology Completes RMB 2 Billion Funding; with Giants Ahead and Pursuers Behind, How Will It Break Through?" — https://cj.sina.cn/articles/view/1650111241/625ab30902001g6xy
7. Science and Technology Innovation Board Daily (reprinted via moomoo) "ShengShu Technology, Valued at Over RMB 12 Billion and Raising RMB 2.6 Billion in Two Months, Plans IPO Possibly in First Half of the Year" — https://www.moomoo.com/hant/news/post/68185217
8. Baidu Baike "AI Comic Drama" — https://baike.baidu.com/item/AI%E6%BC%AB%E5%89%A7/68788906
9. Baidu Baike "Keyframe Animation" — https://baike.baidu.com/item/%E5%85%B3%E9%94%AE%E5%B8%A7%E5%8A%A8%E7%94%BB/10223838
10. 360 Baike "Keyframe" — https://baike.so.com/doc/6737995-32354145.html
11. Measures for Labeling AI-Generated Synthetic Content (issued by the Cyberspace Administration of China and three other departments, CAC Decree No. 2 of 2025) — https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
12. Mandatory national standard GB 45438—2025 "Network Security Technology — Labeling Methods for AI-Generated Synthetic Content" — https://www.tc260.org.cn/upload/2025-03-15/1742009439794081593.pdf
13. Toutiao "The Monetization Logic of AI Comic Drama Can Be Summarized as 'One Foundation, Three Wings'" — https://www.toutiao.com/article/7678962433607664163
14. The Paper "RMB 1,914 to Produce One Episode? Comic Drama Is Still Trapped in Hidden Costs" — https://www.thepaper.cn/newsDetail_forward_33993493
15. Epwk.com "AI Comic Drama Storyboard Design Guide" — https://gonglue.epwk.com/322844.html