Meitu Design Studio / Meitu AI

1. Introduction

1.1. Platform Overview

Meitu Design Studio is Meitu Inc.'s (1357.HK) productivity product built for e-commerce design workflows. Among the six platforms in this group, it is the one with the highest degree of Harness adoption: the Agent Teams upgraded in 2026-06 entrust the entire chain "from product-selection analysis to final-image output" to multiple specialized Agents working in a division of labor, and the company explicitly lists "the creative assets users continuously accumulate within the product can be reused across subsequent tasks" as its third competitive advantage.

By this group's core thesis — the Harness dividing line sits at L3 and L4 — Meitu Design Studio is the only platform in this group that has been confirmed by official documents as having a competitive advantage on both L3 (multi-agent orchestration) and L4 (creative asset reuse) simultaneously. It is also the only platform in this group backed by audited/publicly disclosed financial data, so the causal chain between its Harness investment and business results is the most verifiable.

1.2. Product Line Matrix

In its 2026 interim report, Meitu explicitly divided its business into two lines — lifestyle scenarios and productivity scenarios — and the productivity scenario is the focus of this article.

1.3. 2026 H1 Operating Data

1.3.1. Company Overall (Official, Very High Confidence)

1.3.2. Productivity Applications (Official, Very High Confidence)

1.3.3. Meitu Design Studio Specifics (Official, Very High Confidence)

Data source: Meitu Inc. 2026 interim results announcement (hkexnews) + the Meitu official website; very high confidence.

1.4. Pricing System

Three sets of conflicting figures; must be marked [To be verified]. All of the following come from third-party sources, not from Meitu's official pricing page.

AI outfit-change (virtual fitting) call price:

The two sets of prices differ by about 3×; both are from third-party sites, low confidence. Before selection and budget planning, you must rely on Meitu's official channels (the designkit.com pricing page or sales quotes).

Confirmed free allowance (official FAQ): new users get 20 credits upon registration, and 10 credits per day for daily logins.

2. Glossary

2.1. Common AI Image Terms

2.2. Meitu Design Studio-Specific Terms

2.3. Common Terms for Outfit-Change and Face-Swap

Note: Meitu Design Studio is one of the platforms in this group with the highest productization of the outfit-change (VTON) direction, offering an independent AI fashion-model / AI outfit-change capability and pricing it separately.

Outfit-change error codes (low-confidence source, for reference only): E101 abnormal pose / E102 insufficient lighting / E103 garment segmentation failure.

3. Feature Description

3.1. Full-Product E-commerce Visual Pipeline

Batch processing being a built-in capability is technically critical: most comparable platforms leave batching to the user's own looping calls, whereas Meitu made it a built-in product feature — a sign that L3 orchestration-layer capability has been pushed down into the product layer.

3.2. AI Fashion Model Try-On and Outfit-Change

• Input: flat-lay or dress-form image + select gender / age / ethnicity / body type.
• Output: virtual-model wearing effect, supporting 3:4 / 1:1 / 9:16 ratios.
• Pre/post-processing: smart crop (pre-processing), smart lighting enhancement (post-processing).
• Tested tips (low-confidence source): disabling "smart lighting enhancement" can reduce the probability of distortion by ~30%.
• Pricing: priced separately, with two conflicting sets of figures — ¥0.398/call and ¥1.2/call.

3.3. Multi-Platform Size Adaptation

The same asset is automatically adapted to platform specifications such as Amazon / Shopify / Mercado Libre / TikTok Shop — for example, a main image with a pure-white background RGB 255,255,255.

This feature looks simple, but it is the capability in this group closest to a "business system": platform specifications are encoded into the product rather than left for operators to look up in tables. It is the most straightforward embodiment of Harness value — turning industry rules into default behavior.

3.4. Batch Processing and Commercial Licensing

• Commercial licensing: assets generated on paid tiers are 100% royalty-free and commercially usable (official FAQ).
• Free allowance: new users get 20 credits upon registration, and 10 credits per day for daily logins (official FAQ).
• EXIF metadata compliance: a third-party review claims that images generated by Meitu Design Studio automatically embed the EXIF metadata declaration "AI-generated", making it the most standardized on compliance among domestic tools. This claim comes from a single third-party source, medium confidence, and must be re-verified by actual testing.

4. Platform Architecture

图 4-1｜美图设计室平台架构：模型底座 × Agent 编排 × 即时部署

数据来源：基于本文分析绘制的示意图。

4.1. Model Foundation and Hybrid Orchestration

• Foundation: in-house Meitu MiracleVision model + third-party model / API hybrid orchestration.
• Meitu has explicitly stated that its strategy is to "orchestrate the most suitable models and Skills" — this phrasing is itself Harness thinking: rather than pursuing one model to handle everything, it treats model selection as a schedulable orchestration decision.
• This is conceptually consistent with FLUX.2's "model selection is orchestration" (klein / pro / flex / max routing), but Meitu pushes it down to the "Skill" granularity.

4.2. Agent Layer

Agent Teams (upgraded 2026-06) follows a three-stage structure:

任务拆解（Agent Teams 理解用户意图并拆解）
    ↓
多 Agent 分工（调用最合适的模型与 Skill）
    ↓
精修编辑权交回用户（人仍在环内）

The final step — "returning the fine-editing privilege to the user" — is worth noting: Meitu does not pursue full automation; instead it treats the Agent's output as an editable intermediate state. This is a human-in-the-loop orchestration design, which is more pragmatic than full automation in detail-sensitive scenarios like e-commerce.

4.3. Distribution and Instant Deployment

• Web (bilingual Chinese/English sites) + desktop client + e-commerce plugins.
• An "instant deployment" pipeline for Amazon / Shopify — generated results can enter the sales channel directly, forming a "generate-then-list" closed loop.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Meitu's context is composed of four kinds of elements: product-selection info + product image + platform specifications + brand tone.

Among these, "platform specifications" being explicitly included as a context element is the L1 design in this group with the most business specificity — it means the model "sees" not only "what to draw" but also "where the result will be used and what hard constraints apply." Rules such as Amazon main images requiring a pure-white background RGB 255,255,255 are, in Meitu's case, system constraints encoded into the generation conditions rather than manual check items.

Agent Teams takes on intent understanding and task decomposition, turning a user's single natural-language request into a structured assembly of context.

5.3. L2 Tools and Execution Layer

• Tool set: matting, background swap, erasing, upscaling, try-on, image set, video, multi-platform adaptation.
• Orchestration strategy: in-house MiracleVision model + third-party APIs in hybrid orchestration; Meitu has explicitly stated it "orchestrates the most suitable models and Skills."
• Granularity: down to the Skill level, not just the model level — this aligns most closely with the L2 definition of "tool registration, sandboxing, parallel/serial scheduling" in the parameter card.
• Built-in batching: whole-batch processing is a built-in capability, requiring no external loop.

5.4. L3 Orchestration and Control Layer

Meitu Design Studio's Agent Teams is the form closest to a third-generation Harness among the six platforms in this group.

The L3 comparison across the six platforms in this group:

The difference between Meitu and ComfyUI (the strongest L3 reference in this group) is that ComfyUI's orchestration artifact is a version-controllable JSON graph, while Meitu's orchestration artifact currently has no public statement about exportability or version control. Thus Meitu is strongest in "orchestration capability" but still lags ComfyUI in the "manageability of orchestration artifacts."

Delivery closed loop: generation → multi-platform size adaptation → instant deployment, forming a "generate-then-list" flow. It is the only platform in this group that connects the end point of orchestration to a sales channel rather than a file system or gallery.

5.5. L4 Memory and State Layer

In its 2026 interim report, Meitu explicitly lists "creative asset reuse" as the third competitive advantage, stated verbatim as: the creative assets users continuously accumulate within the product can be reused across subsequent tasks without starting from scratch.

This aligns with Kling's "subject creation (Element)" as an L4 asset-ization approach, but with a key difference:

Putting L4 into the competitive-advantage list of its financial results is the strongest signal in this group that "L4 has been commercially validated." It means: asset reuse is not just a feature but a driver of retention and consumption depth — consistent with Meitu's disclosure that "several thousand users' annual consumption reaches about 10× the blended ARPPU, and that cohort has grown about 6× since 2025-12."

Still missing: no public version management of workflow artifacts, no model-version-pinning endpoints, and no public explanation of cross-platform asset export.

5.6. L5 Evaluation and Observability Layer

Meitu's L5 is the most "business-oriented" in this group: using consumption depth to measure value hit rate.

These metrics are not image-quality metrics (SSIM / FID / LPIPS) but business-health metrics. Their advantage is that they align directly with business results; their disadvantage is that they cannot tell whether a single image is good — so Meitu performs "system-level observation" at the evaluation layer, while "image-level evaluation" remains absent.

5.7. L6 Governance and Security Layer

• EXIF metadata AI declaration: a third-party review claims that images generated by Meitu Design Studio automatically embed the EXIF metadata declaration "AI-generated", making it the most standardized on compliance among domestic tools. Single third-party source, medium confidence, must be re-verified by actual testing.
• Commercial licensing: assets generated on paid tiers are 100% royalty-free and commercially usable (official FAQ).
• Outfit-change image boundary: AI outfit-change produces a virtual model figure, not involving any real natural person's identity, so portrait-right risk is significantly lower than real-person face-swap. But if a user uploads a real person's photo for garment replacement, it falls within the scope of Article 17 of the 《Deep Synthesis Regulations》.
• 《Labeling Measures》 compliance: if the EXIF declaration holds, it can correspond to the Article 5 requirement of implicit labeling (file metadata); details of the explicit-labeling implementation are not publicly described, [To be filled].
• Organizational governance reference: the internal studio mechanism (each studio gets an initial budget of ≈ ¥10 million and is stopped if it does not reach commercialization milestones) is organizational-level innovation governance that can serve as reference material for the "governance" chapter — it institutionalizes "trial-and-error cost" in a terminable way.

5.8. Maturity Assessment

Meitu Design Studio is a "strong model + strongest Harness (in this group)" form. On all four layers — L2 (hybrid orchestration of models and Skills), L3 (Agent Teams), L4 (creative asset reuse), and L5 (business-metric observability) — it has the highest maturity in this group, and each layer is corroborated by public financial results or official documents rather than third-party speculation. Its only notable weakness is L6 — the EXIF declaration must be re-verified by actual testing, and the explicit-labeling implementation details are not publicly disclosed.

Judged against this group's core thesis: Meitu Design Studio is positive proof of the judgment that "the Harness dividing line sits at L3/L4" — its business results (ARR of RMB 620 million, 8× compute-point consumption growth in 8 months, 6× high-value-user growth) are highly correlated with its L3/L4 investment.

6. Actual Cases

6.1. Official Quantified Results of Meitu Design Studio Agent Teams

All of the following are data publicly disclosed in Meitu Inc.'s 2026 interim results announcement (very high confidence):

6.2. Kaipai and Vmake Labs

6.3. MVLAND and RoboNeo

6.4. Seller Scenarios in Third-Party Reviews

Third-party reviews claim: processing quality in the apparel / beauty / portrait categories ranks in the domestic first tier; "natural skin tone, high light-fusion quality, and a lower AI feel than most competitors"; and on Amazon, an image set of about 7,500 beans can generate 72 times.

This source is a third-party review with low confidence; when cited it must be labeled as a community / third-party basis and must not be used as official data.

7. Summary

7.1. Strengths

1. L3 orchestration is the strongest in this group: Agent Teams multi-agent collaboration covers the full chain from product-selection analysis to final-image output.
2. L4 asset reuse officially confirmed as a competitive advantage: the interim report explicitly lists "creative assets can be reused across subsequent tasks" as the third competitive advantage, making it the only L4 design in this group backed by official documents.
3. L5 observability is directly tied to business results: compute-point consumption QoQ, ARR, and ARPPU multiple distribution form a verifiable value closed loop.
4. The end point of delivery connects to a sales channel: multi-platform size adaptation + instant deployment achieves "generate-then-list".
5. Batch processing is built in: no external looping calls required.
6. Highest data credibility: the only platform in this group backed by publicly disclosed financial results (RMB 2.21 billion revenue, 18.44 million paid subscriptions, 33 million productivity MAU, ARR of RMB 620 million).
7. Outfit-change does not involve real natural-person identity: the virtual-model path carries significantly lower portrait-right risk than real-person face-swap.

7.2. Limitations and Applicability Boundaries

Applicability boundaries: best suited to scaled visual production for e-commerce sellers and agency operations teams (product images, model images, detail pages, multi-platform adaptation); not suited to general concept creation, cinematic-quality imagery, or scenarios requiring offline, privately deployed setups.

7.3. Selection Recommendations

1. Evaluate by the value of "instant deployment" rather than by per-image cost: Meitu's difference lies in results that can be listed directly, saving operational hours rather than the unit price of images.
2. Pricing must be confirmed with the official source: both the membership price and the outfit-change call price have multiple third-party sets of figures, with gaps of up to 3×; obtain an official quote before signing.
3. Run a small-batch trial of the outfit-change scenario first: error codes E101 / E102 / E103 (pose / lighting / segmentation) show that input quality directly determines success or failure; it is recommended to first run 50–100 images to evaluate the pass rate.
4. Build your own explicit-labeling step: even if the EXIF declaration holds, the explicit label required by Article 4 of China's 《Labeling Measures》(a prominent prompt label in an appropriate position on the image) still needs to be injected into the export pipeline on the business side.
5. Include compute-point consumption in observability: if you adopt Agent-ization capabilities, you can follow Meitu's experience metric (QoQ growth in compute-point consumption) to measure adoption depth, rather than looking only at account counts.

7.4. Compliance Notes

• 《Measures for Labeling AI-Generated and Synthetic Content》(国信办通字〔2025〕2 号) Article 4: when providing download, copy, or export functions, you shall ensure the file contains a qualified explicit label (for images, "add a prominent prompt label in an appropriate position"); Article 5: you shall add an implicit label to the file metadata. If the EXIF declaration holds it can correspond to Article 5; the explicit label still must be built on your own.
• Article 17 of the 《Provisions on the Administration of Deep Synthesis in Internet Information Services》: deep-synthesis services such as face replacement and face generation that may cause public confusion or misidentification shall be prominently labeled. If AI outfit-change is used for real-person image replacement, this article applies.
• Article 1019 of the 《Civil Code of the People's Republic of China》: no one may infringe another's portrait right through falsification using information-technology means or otherwise. The virtual-model path does not involve any real natural person, but if a real person's photo is uploaded for garment replacement, the authorization of the portrait right holder must be obtained.
• Beijing Internet Court effective judgment of 2026-03: identifiability + burden-of-proof shift. Once e-commerce material involves a real person's image, the generation parameters and the authorization chain must be preserved.

Information-Gap Declaration

1. Meitu Design Studio membership pricing: three conflicting sets of figures (¥30 / from ¥88; ¥49 including membership; bean-based metering), all from third-party sources.
2. AI outfit-change call price: two sets of figures (¥0.398/call basic, ¥0.15/call enterprise vs ¥1.2/call, 1,000-call pack ¥800), all from third-party sites.
3. EXIF "AI-generated" metadata declaration: a single third-party review source, recommended to be re-verified by actual testing.
4. Specific monthly "bean" allowance and exchange ratio: only third parties mention "the premium tier includes up to 7,500 beans per month".
5. Technical implementation details of Agent Teams (number of Agents, division-of-labor method, scheduling mechanism, whether customizable): the interim report gives only a qualitative description, [To be filled].
6. Specific form and retention period of creative assets: the interim report only states "creative assets can be reused", with no detailed rules, [To be filled].
7. Outfit-change error codes E101 / E102 / E103: low-confidence source.
8. Details of explicit-label implementation under China's 《Labeling Measures》: no official public explanation, [To be filled].
9. Whether orchestration artifacts can be exported / version-controlled: no public explanation, [To be filled].
10. Officially published single-merchant customer cases (non-financial-report basis): none found. [No results]

8. References

1. Meitu Inc. 2026 interim results announcement — Meitu Inc. (1357.HK), 2026-08-26. https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0826/2026082600321_c.pdf
2. Meitu Inc. 2026 H1 financial report — Meitu Inc. official website, 2026-08-26. https://preview-www.meitu.com/zh/media/439
3. Designkit official website (Meitu Design Studio's overseas site) — Meitu Inc. https://www.designkit.com/
4. Guangming Online ·《From "helping you create" to "helping you do business": Meitu AI productivity enters its value-realization period》 — 2026-08-26. https://tech.gmw.cn/2026-08/26/content_38966018.htm
5. China Daily ·《Behind Meitu's 2026 interim results: a three-layer restructuring of people, money, and product》 — 2026-08-27. https://cn.chinadaily.com.cn/a/202608/27/WS6a8fde00e4b09a165c78658e.html
6. 《Measures for Labeling AI-Generated and Synthetic Content》(国信办通字〔2025〕2 号) — CAC et al., effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
7. Interpretation of the 《Measures for Labeling AI-Generated and Synthetic Content》(explicit labels distinguished by medium) — Gov.cn / Xinhua, 2025-03-16. https://www.gov.cn/zhengce/202503/content_7014281.htm
8. Article 17 of the 《Provisions on the Administration of Deep Synthesis in Internet Information Services》 — CAC et al., 2022. (Provision quoted in the body; no official link)
9. Article 1019 of the 《Civil Code of the People's Republic of China》 — NPC, 2020. (Provision quoted in the body; no official link)
10. Economic Information Daily ·《Technology is not an infringement "shield"; the court ruled this way on AI "face theft"》 — Xinhua 《Economic Information Daily》, 2026-04-17. http://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm
11. Tianji Net ·《A guide to Amazon-seller AI image-generation tools: a review of 10 tools》(includes Meitu Design Studio pricing, third-party low confidence) — https://news.yesky.com/hotnews/263/371763.shtml
12. Dashu Kuajing · Meitu Design Studio AI outfit-change (fees / error codes / applicable categories, third-party low confidence) — https://applet.10100.com/encyclopedia/explain/81219759
13. Official ComfyUI workflow 「Virtual Character Try-On - 4-in-1」(a comparison baseline for version-controllable workflows) — Comfy Org. https://comfy.org/zh/workflows/templates_rob_fashion_shoot_vton-4in1.app/