Jimeng AI (ByteDance)

1. Introduction

1.1. Platform Overview

Jimeng AI was incubated by ByteDance's Jianying (CapCut) team, and Baidu Baike lists its operating entity as Shenzhen Lianmeng Technology Co., Ltd. Within this platform group it is the domestic sample where the "model—product—cloud API" three-layer pipeline is the most thoroughly connected: at the base are ByteDance Seed team's Seedream (image) and Seedance (video) models; in the middle are ecosystem products such as Jianying / CapCut / Doubao / Xiaoyunque; and externally it provides enterprise APIs through Volcano Engine.

In the six-layer AI Harness capability model, Jimeng's positioning can be summarized as: heavy investment in L1 context engineering (multi-image reference + visual signals + web retrieval + chain-of-thought), while the L6 governance layer has been tested by regulators and exposed for its shortcomings. It is the only platform in this group publicly recorded as being penalized for failing to implement the requirements on marking AI-generated synthetic content — a fact that makes it a key sample for studying "how the L6 governance layer shifts from after-the-fact remediation to built-in mechanisms."

1.2. Development Milestones

1.3. Pricing System

1.3.1. C-side Memberships (two conflicting sets of figures)

Both sets of figures come from third-party WeChat screenshots and self-media reviews, neither from official channels; this report presents them side by side and uniformly marks them [To be verified]. For selection decisions, use the real-time price in the App.

Converted according to Figures A, 1,080 credits roughly correspond to 4,320 images (image generation is on the order of about 4 credits/image); video consumption is significantly higher. This conversion is an estimate.

1.3.2. B-side Volcano Engine API (official, last updated 2026-03-31)

This set of unit prices is among the highest-confidence pricing data in this group and can be used directly for cost estimation. Comparison within the group: Tongyi Wanxiang wan2.7-image is 0.2 yuan/image, Qwen-series image editing 0.14 yuan/image, FLUX.2 [klein] 4B is $0.014 + $0.001/MP.

2. Glossary

2.1. General AI Image Terminology

2.2. Jimeng- and Seedream-Specific Terminology

2.3. General Terminology for Try-On and Face Swap

Note: In its "eight core capabilities", Jimeng / Seedream officially lists virtual try-on as a typical scenario of "multi-image reference", but does not provide a standalone try-on product line or masked face-swap capability.

3. Feature Description

3.1. Generation and Editing Capabilities

3.2. Smart Canvas

The Smart Canvas is Jimeng's most important productized expression of L1 context engineering: it organizes "multiple reference images + layers + locally painted regions" into a visual context workspace. Unlike Midjourney's "parameter string" and FLUX.2's "JSON fields", Jimeng chose a spatial (canvas) way of organizing context — whatever the user places on the canvas is what the model sees.

Advantages: intuitive, zero learning cost; disadvantages: the context is hard to serialize and version, and cannot be entered into Git for diff and regression like ComfyUI's JSON graphs.

3.3. Multimodal and Narrative Capabilities

• Multimodal input: Seedance 2.0 supports mixed input of four modalities: image, video, audio, and text.
• Narrative orchestration: storyboarding + Xiaozhangyu Octo (collaborative narrative creation tool) + first-frame / last-frame constraints for video.
• Film-style extension: launched the "Jimeng Film Set" label on 2026-08-26.

3.4. Commercial and Asset Governance

• Commercial licensing: membership tiers include the "remove brand watermark from generated works" entitlement (from screenshots, medium confidence).
• Real-person asset restriction: since the 2026-02 version, the use of real-person assets has been restrictedlinked, and a digital human identity verification mechanism was introduced (per Baidu Baike; specific flow [To be filled]).

4. Platform Architecture

图 4-1｜即梦 AI 平台架构：从 Seed 模型底座到火山引擎云 API

数据来源：基于本文分析绘制的示意图。

4.1. Model Base

Seedream 5.0's eight core capabilities (official technical page): precise editing, flexible reference, visual signal control, in-context reasoning generation, multi-image reference, multi-image output, advanced text rendering, adaptive aspect ratio and 4K.

4.2. Distribution and Ecosystem

• C side: Web (jimeng.jianying.com) + iOS / Android + mini program, directly connected to the Jianying ecosystem.
• Ecosystem embedding: Jianying, CapCut, Xiaoyunque, Doubao.
• B side: Volcano Engine API, using an asynchronous task model + concurrency purchase (500 yuan/day/concurrency, 10,000 yuan/month/concurrency).

4.3. Inference Optimization

Per the official technical page: the Seedream series uses adversarial distillation for stable few-step inference + 4/8-bit mixed-quantization offline smoothing + speculative decoding to reduce latency; the vendor claims DiT image generation is more than 10x faster than Seedream 3.0 (vendor's own claim, no third-party reproduction seen).

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Within this platform group, Jimeng belongs to the camp with the heaviest L1 investment; its context is orchestrated uniformly across five categories of signals:

1. Text prompts: Chinese semantic understanding ranks in the domestic first tier; no English prompt tricks needed.
2. Multi-image reference: official figures of "a dozen or so" (third-party API figures of 1–10), able to extract character features, scene style, and object structure for organic fusion.
3. Structural signals: Canny / Depth / Mask natively built in, no external ControlNet model needed — this eliminates tool dependency at the engineering level and is a key difference from Midjourney.
4. External knowledge: Seedream 5.0 first supports real-time web retrieval (RAG), bringing external facts into the generation context.
5. Reasoning chain: equipped with CoT chain-of-thought reasoning, supporting multi-step logical reasoning and integration of web knowledge.

Officially, virtual try-on is clearly listed as a typical scenario of multi-image reference — tasks such as "maintaining garment scale and physical coherence while fusing character features" are classified under context engineering rather than as a standalone product capability.

5.3. L2 Tools and Execution Layer

• Tool set: generation, repainting, canvas expansion, removal, cutout, upscaling, dubbing, video.
• Exposure form: canvas buttons + Volcano Engine API (asynchronous tasks + concurrency expansion).
• No open tool registration mechanism: third parties cannot register new tools; Function Calling or MCP is not supported.
• Good cost predictability: the B side clearly lists prices per call / per generated image and can be estimated precisely; the C side is priced by credits but has conflicting pricing sets.

5.4. L3 Orchestration and Control Layer

Jimeng's orchestration is lightweight orchestration, sitting between Midjourney's "no orchestration" and Meitu Design Studio's Agent Teams "multi-agent orchestration":

Gap: no DAG, no sub-agent dispatch, no interruption and resume, no workflow artifact version control.

5.5. L4 Memory and State Layer

• Cloud asset library / creation history / template library: provides basic state persistence.
• Digital human identity verification (2026-02): consolidates the "person" into a reusable asset — a noteworthy L4 design in this group, following the same line of thinking as Miaoya Camera's "digital avatar" and Kling's "subject creation (Element)", but Jimeng ties it to asset governance (restricting real-person asset use), making it the only design in this group that explicitly couples L4 with L6.
• Gap: no engineered Checkpoint, artifact version management, or model-version locking endpoints. The Smart Canvas context is spatial, making it hard to serialize and version.

5.6. L5 Evaluation and Observation Layer

• Officially claims "leading in comprehensive benchmarks, No.1 in global Elo for text-to-video and image-to-video"; self-media cites Elo 1269 / 1351. This figure is a value at a specific event time point, should be labeled as vendor's own claim or a third-party leaderboard time-point value, and is not recommended as a stable fact to cite.
• No official Eval Set, Golden Dataset, or regression set disclosed.
• Observation basis: credits consumed on the C side, call count and concurrency on the B side. No pre-generation cost estimation tool is seen (compare Leonardo.ai's Pricing Calculator endpoint).

5.7. L6 Governance and Safety Layer

This is the layer with the greatest sample value in this group, because it has already been tested by regulators:

Fact: on 2026-04-28, the Jimeng AI website was penalized by cyberspace regulators for failing to effectively implement the requirements on marking AI-generated synthetic content. (Source: Baidu Baike's "Jimeng AI" entry; medium–high confidence; recommend cross-checking with the official notice.)

Against the requirements of the Measures for Marking AI-Generated Synthetic Content (Guoxinban Tongzi [2025] No. 2):

Since 2026-02, Jimeng has proactively restricted the use of real-person assets and introduced digital human identity verification, showing that its governance awareness is not behind; yet "marking" — this most basic and most rigid requirement — still shows a gap in implementation. This confirms the group's core conclusion: L6 cannot be achieved by policy statements; it must be achieved through mechanisms and engineering pipelines — the label must be injected automatically in the download / export chain, not left to operational configuration.

5.8. Maturity Assessment

Jimeng takes the form of "strong model + medium Harness + governance layer already tested and remediated". Its L1 investment reaches the group's first tier (shared with Seedream's model capability), L3 has lightweight orchestration but no workflow artifacts, L4 has an asset library but no version management, L5 is not engineered, and L6 has a confirmed historical gap. It demonstrates that: neither strong model capability nor a good ecosystem position can replace the engineering implementation of the labeling chain.

6. Real-World Cases

6.1. Sanxingdui: Revelations of the Future

• Time: 2024-07.
• Content: China's first AIGC-generated sequential-narrative sci-fi short drama series launched on Douyin, with Jimeng AI as chief AI technology supporter.
• Technical output: the collaboration improved video-generation capabilities, including 24/30/60 fps frame interpolation, 2x upscaling, and control over horizontal and vertical camera movement and direction/magnitude.
• Nature: a real collaboration case, but no quantified performance data disclosed (view counts, conversion rates, etc.).
• Confidence: high (Baidu Baike "Jimeng AI" entry).

6.2. 38th Hundred Flowers Awards AIGC Recommendation Unit

• Time: 2026-08-08.
• Content: Jimeng AI participated in the supporting activities as exclusive AIGC technology partner.
• Nature: brand partnership, no quantified performance data. Medium–high confidence.

6.3. Jimeng Film Set and the Film & TV Content Label

• Time: 2026-08-26.
• Content: launched the film & TV content label "Jimeng Film Set", extending toward film industrialization.
• Nature: a product move, not a customer case. Medium confidence.

6.4. Regulatory Penalty Case (negative example)

• Time: 2026-04-28.
• Content: the Jimeng AI website was penalized by cyberspace regulators for failing to effectively implement the requirements on marking AI-generated synthetic content.
• Nature: the only real regulatory case in this group that can be written into the document, and direct evidence for the group's argument that "the L6 governance layer must be implemented through engineering mechanisms". Medium–high confidence; recommend cross-checking with the official notice.

7. Summary

7.1. Strengths

1. Chinese semantic understanding ranks in the domestic first tier, requiring no English prompt tricks, so teams face the lowest onboarding cost.
2. Visual signals natively built in (Canny / Depth / Mask), no external ControlNet needed, fewer engineering dependencies.
3. Highly transparent B-side pricing: officially listed by Volcano Engine (0.2 yuan/image, 0.22 yuan/image, 0.2 yuan/call inpainting), enabling direct cost estimation — among the highest-confidence pricing data in this group.
4. Deep ecosystem embedding: Jianying / CapCut / Xiaoyunque / Doubao form natural distribution, so model capabilities can be "released the same day and reach hundreds of millions of users".
5. Web retrieval + CoT: brings external facts and reasoning chains into the generation context, a substantive extension of L1.

7.2. Limitations and Applicability Boundaries

Applicability boundary: suited to Chinese content-creation teams, short-video and short-drama production, and batch generation of e-commerce marketing images (via API); not suited to enterprise-grade pipelines that require strong compliance auditing, reproducible regression, and cross-system orchestration, unless a labeling and version-management layer is built in-house.

7.3. Selection Recommendations

1. Procure by API rather than membership: Volcano Engine pricing is the official baseline and can be estimated; the two conflicting C-side membership price sets should not be used for budgeting.
2. Build a labeling chain in-house: after calling the Jimeng API to generate, before distribution, the business side should uniformly inject the explicit label and metadata implicit label rather than relying on the platform's default behavior — an engineering conclusion drawn directly from the 2026-04-28 penalty event.
3. Supplement L4/L5: use an external system to record the "prompt + reference-image hash + model version + parameters" four-tuple as the minimum reproducible unit and evidentiary material.
4. Real-person assets: be sure to complete the digital human identity verification process first (specific flow [To be filled]), and retain authorization evidence.

7.4. Compliance Notes

• The Measures for Marking AI-Generated Synthetic Content (Guoxinban Tongzi [2025] No. 2), effective 2025-09-01: Article 4 explicit-label obligation, Article 5 implicit-label obligation, Article 10 prohibition on maliciously deleting, altering, forging, or concealing labels.
• Civil Code of the People's Republic of China, Article 1019: must not infringe others' right of portraiture through forgery or other technical means; the fair-use circumstances in Article 1020 do not include commercial face-swapping.
• Provisions on the Administration of Deep Synthesis of Internet Information Services, Article 17: deep-synthesis services such as face replacement and face generation that may cause public confusion or misidentification shall carry a prominent label.
• Effective judgment of Beijing Internet Court in 2026-03: shift of burden of proof — a defendant claiming "an AI coincidentally looks like someone" must reproduce the creation process; failing to reproduce results in the adverse consequence of failed proof. When using tools such as Jimeng that keep no version records, proactively save the generation parameters and reference images for evidence purposes.

Information Gap Statement

1. C-side membership pricing: two conflicting sets of figures (¥69/¥299/¥499 and ¥79/¥239/¥649), both from third-party sources. [To be verified; use the real-time price in the App]
2. The specific flow of the digital human identity verification mechanism: only mentioned in Baike, no details. [To be filled]
3. The original text of the official notice of the 2026-04-28 penalty event: the current source is the Baidu Baike entry.
4. Seed locking, CFG exposure, and model-version locking endpoints: no official documentation found. [To be filled]
5. Implementation details of explicit and implicit labels under China's Marking Measures: no official public notes seen. [To be filled]
6. Official merchant cases with quantified performance data: none found. The Sanxingdui and Hundred Flowers Award collaborations are real cases but contain no performance data. [No quantified data]
7. Seedream 5.0 "more than 10x faster than 3.0": vendor's own claim, no third-party reproduction seen.
8. Elo 1269 / 1351: event time-point values cited by self-media, not recommended as stable facts to cite.

8. References

1. Volcano Engine · Jimeng AI - image generation billing notes — Volcano Engine, last updated 2026-03-31. https://www.volcengine.com/docs/85621/1544714
2. Jimeng AI official website — ByteDance. https://jimeng.jianying.com/
3. Seedream 5.0 technical page (eight capabilities / unified architecture) — ByteDance Seed team. http://seedream4.org/seedream-4
4. Baidu Baike · Jimeng AI (includes penalty event and product history, secondary source) — https://baike.baidu.com/item/%E5%8D%B3%E6%A2%A6App/67386767
5. Baidu Baike · Seedream (includes release dates and team changes, secondary source) — https://baike.baidu.com/item/Seedream/67390954
6. The Measures for Marking AI-Generated Synthetic Content (Guoxinban Tongzi [2025] No. 2) — Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security, National Radio and Television Administration, published 2025-03-14, effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
7. Interpretation of the Measures for Marking AI-Generated Synthetic Content — gov.cn / Xinhua News Agency, 2025-03-16. https://www.gov.cn/zhengce/202503/content_7014281.htm
8. Provisions on the Administration of Deep Synthesis of Internet Information Services, Articles 16 and 17 — Cyberspace Administration of China et al., 2022. (Clauses cited in the text; no official link)
9. Civil Code of the People's Republic of China, Articles 1018, 1019, and 1020 — National People's Congress, 2020. (Clauses cited in the text; no official link)
10. Economic Information Daily · "Technology is not a 'shield' for infringement; how courts determined AI 'face theft'" — Xinhua News Agency's Economic Information Daily, 2026-04-17. http://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm
11. Alibaba Cloud Bailian · wan2.7-image model documentation (benchmark for horizontal comparison in the same price band) — Alibaba Cloud. https://help.aliyun.com/zh/model-studio/wan2-7-image
12. Black Forest Labs · FLUX.2 Overview (benchmark for structured prompts and fixed-snapshot endpoints) — BFL, 2026. https://docs.bfl.ai/flux_2/flux2_overview