Miaoya Camera (妙鸭相机)

1. Introduction

1.1 Platform Overview and Positioning

Miaoya Camera was the first phenomenon-level product of the 2023 AI portrait (AI 写真) track: users upload a batch of photos of themselves, pay to create a "digital avatar", and then batch-generate portrait outputs in different styles from that avatar. Its operating entity is Weixu Network Technology (Shanghai) Co., Ltd., incubated by a founding team in which Alibaba's Digital Media & Entertainment Group (DMEG) participated, and later merged into "Shenli Shijie", a wholly-owned company of Alibaba DMEG; its technical foundation is the "Tixiang" model developed in-house by Alibaba DMEG, which analyzes multi-angle photos to generate a 3D face model and supports dynamic expression simulation.

In the six-layer capability model of AI Harness, Miaoya is an extreme single-point sample: at L1 (turning the "person" itself into context) it made the most radical attempt of the time, at L4 (digital avatar as a persistent identity asset) it created this group's first mass-market "identity asset product", yet L2, L3, and L5 were nearly blank, and L6 suffered a public trust incident early on. It ultimately did not survive — the team was formally disbanded at the end of September 2025, and the product is kept running at minimum-level operation only, with no further feature upgrades or promotion.

This outcome makes Miaoya the only negative sample with a complete lifecycle in this group's research. Its value lies not in "what it did wrong", but in proving a proposition: given that generation quality is already good enough, what determines a product's survival is not the model but the Harness that carries the model.

1.2 Development History and Lifecycle

It should be noted that: the authoritative source for the fact of the team's disbandment is only the wording on a third-party tools website; no corroboration from tech media reports or business/judicial information has been located, marked [To be verified]; the current App version number also has two conflicting figures (v4.6.14 updated on 2026-08-18, and v4.3.1 updated on 2026-04-08), marked [To be filled]. All analyses in this section that are based on the fact of "disbandment" presume that the fact holds; if later review overturns it, the related attribution conclusions need to be re-evaluated.

1.3 Pricing and Open Form

The pricing above is point-in-time data from 2023, from consistent wording across multiple third-party sources, but none are archived official original pages, marked [To be verified]. Miaoya's pricing structure itself already exposed the problem with its business model: a one-time ¥9.9 buyout of one avatar + 10 outputs is typical "loss-leader pricing", lacking subsequent consumption scenarios and subscription anchors — a stark contrast with Kling's "Elements" (subject creation), another identity-type product, which limits 30 / 50 / 150 / 500 subjects by membership tier.

2. Glossary

2.1 General Direction Terms

2.2 Miaoya-Specific Terms

3. Feature Description

3.1 Core Functional Chain

Miaoya's functional chain is extremely short, with only three steps:

1. Upload and Training: the user uploads ≥20 multi-angle, multi-expression, multi-lighting half-body photos of themselves, and the system builds a digital avatar (requiring a queue, from several hours to a dozen-plus hours at peak);
2. Template Application: choose from 30+ style templates, and the avatar drives generation of portrait outputs in the corresponding style;
3. Retouching and Download: retouch and HD-download the selected output (consuming Diamonds), or use AI photo retouching (added in the 2025-09 version).

The engineering feature of this chain is: every interaction is an atomic operation. User upload → wait → choose template → generate output; among these four steps there is no intermediate state to intervene in, no parameter to tune, and no batch-processing entry. When an output's quality is unsatisfactory, the only things a user can do are "choose a template again" or "fine-tune with Make It More Like Me" — unlike Kling O1, which can iterate and correct in the same session with natural language, or ComfyUI, which can go back to some node and rerun.

3.2 Template and Style System

The design of the template system exposes a key trade-off: Miaoya put all of the "controllability" on the platform side and all of the "investment" on the user side. Users paid 20 photos and hours of queuing, yet got zero control over the generation process. This asymmetry is masked by "novelty" in a product's early days, but once the novelty fades it inevitably turns into a retention problem.

3.3 B-Side and Open Capabilities

Miaoya once attempted to extend toward the B side: by opening an API, it cooperated with platforms such as Ctrip to launch enterprise-level services like "AI Travel Photography". This information has low confidence and is marked [To be verified] — no official announcement, cooperation press release, or Ctrip-side product page has been found to corroborate it.

From a Harness perspective, even if that cooperation were real, Miaoya lacked the infrastructure to support B-side engineering integration: no public API documentation, no concurrency/quota contracts, no callback mechanism, no error-code specification. By contrast, enterprise-oriented image platforms of the same period (such as Leonardo.ai) had already explicitly documented the three capacity constraints of rate limits, concurrency caps, and queue depth, and provided Webhook and idempotency guidance. This contrast is the watershed of "whether a Harness exists".

4. Platform Architecture

图 4-1｜妙鸭相机训练式身份管线：从照片到写真成片

数据来源：基于本文分析绘制的示意图。

4.1 Training-Based Identity Pipeline

Miaoya adopts a training-based (per-user fine-tune) technical route rather than zero-shot adaptation:

≥20 张多角度照片 → 人脸检测与对齐 → 提香模型 per-user 训练 → 数字分身（人脸 / 形象模型）
                                                                    ↓
                                        风格模板（提示词 + 构图 + 光影）→ 合成 → 成片 → 精修 / 高清下载

This is a fundamental difference from the "zero-shot identity injection" route of InstantID, PuLID, etc., and is the root cause of its persistently high cost and waiting time:

The training-based route was a reasonable choice in 2023 — zero-shot approaches were not yet mature; but Miaoya did not complete the route migration after zero-shot approaches matured in 2024—2025, instead cementing its cost structure. This is a typical case of technical debt turning into commercial failure.

4.2 Distribution Form

• The primary entry is the Mini Program (WeChat / Alipay), with the App as auxiliary and the official site as a brand showcase only. The mini-program-first strategy brought extremely low acquisition costs and very fast dissemination in 2023, but the price was capabilities constrained by the host environment (file export, background tasks, and compute scheduling are all limited).
• Cross-platform account interoperability (iOS / Android / WeChat / Alipay), with avatars and outputs stored in the cloud.
• No open API (any B-side cooperation, if it existed, was bespoke per case, not a platformized capability).

4.3 Comparison with the Zero-Shot Identity-Injection Route

From a Harness standpoint, the difference between the two routes can be summed up in one sentence: the training-based route turns the identity asset into a "private state within the product", while the zero-shot route turns the identity asset into a "context that can circulate across pipelines". The former looks stronger at L4 (a dedicated model) but is in fact weaker — because it cannot be migrated, combined, or audited; the latter looks lighter, yet gains true asset liquidity because "the identity adapter can be loaded by anything".

5. Harness Design

5.1 Six-Layer Capability Overview

5.2 L1 Context Engineering Layer

At L1, Miaoya made a design that was highly imaginative for its time: compressing the "person" itself into context. Traditional text-to-image context is the prompt; Miaoya's context is "a trained person". From an engineering standpoint, this changes high-dimensional identity information (bone structure, facial-proportion, skin tone, expression habits) from "needing to be re-described for every generation" into "injected once, reused long-term".

The problem with this design is that it is one-shot and non-editable:

• Users cannot adjust "which features are remembered and which are ignored" (contrast Kling's Identity Map, which decouples "stable identity markers" from "transient attributes (expression / outfit / lighting)");
• Users cannot add context for new scenarios (contrast FLUX.2, which supports carrying up to 8—10 reference images with a task);
• The only tuning knob is "Make It More Like Me" — a vague slider with no dimension and no feedback loop.

In other words, Miaoya's L1 is a "write-once, read-only" context, whereas the third-generation Harness requires context to be "assemblable, trimmable, and priority-orderable".

5.3 L2 Tools and Execution Layer

The toolset is extremely thin: template selection, crop, color, defect repair, AI photo retouching (added 2025-09). There is no external tool-registration mechanism, no scripting entry point, no batch processing, and no tool contract of any kind.

For a C-end portrait product, few tools is not fatal; what is fatal is: few tools + no orchestration = every user need must be preset by the platform. Once the 30+ templates are used up, the product loses any new source of value supply.

5.4 L3 Orchestration and Control Layer

This is Miaoya's weakest layer and the direct cause of its user-experience collapse:

"Queue-based rendering" is the most visible product of this layer's failure: it essentially passes compute scarcity directly on to users, without shaping load, degrading, async-ing, or managing expectations through the orchestration layer.

5.5 L4 Memory and State Layer

The digital avatar is Miaoya's only persistent asset and the whole of its value — this sentence is both its product positioning and its death sentence.

Identity Preservation is the shared L4 core challenge of the three product families — outfit-changing, face-swapping, and digital avatars: its technical essence is persisting "identity features" as cross-session state. Whoever makes L4 solid is who keeps users — because users' investment (photos, time, accumulated aesthetic preferences) all attaches to this layer, forming switching costs.

But Miaoya's L4 is a one-way locked asset:

1. Not migratable: the avatar exists only inside the Miaoya product; users cannot export the model or reuse it elsewhere. This means all of a user's investment in Miaoya is one-off sunk cost, not a portable asset.
2. Not combinable: the avatar can only combine with platform-preset templates, not with the user's own reference images, brand assets, or scene constraints.
3. Not versionable: there is no concept of "avatar v1 / v2"; users cannot go back to a more satisfactory identity version.
4. Dies with the product: once the team disbanded and the product stopped operating, users' identity assets disappeared along with it.

Contrast the same track's practitioners who got L4 right: Kling made "Elements" (subject creation) reusable subjects metered by tier (30 / 50 / 150 / 500); Runway made Brand Kits (up to 3) and Custom Voice Clones (up to 3) first-class assets; FLUX.2's Base variant is designed specifically for LoRA fine-tuning, so identity and brand assets are held by users themselves as weights; the ComfyUI ecosystem puts both workflow JSON graphs and LoRA weights under version control. What they share: ownership and portability of assets rest on the user side. Miaoya is exactly the opposite — it kept users' strongest asset on its own servers while failing to offer enough long-term value to make users stay.

5.6 L5 Evaluation and Observability Layer

Miaoya has no evaluation layer in the engineering sense. The only evaluation entry point in the product is "Make It More Like Me" — relying on users' subjective judgment through trial and error, with no objective metrics, no regression set, and no A/B mechanism.

For identity-type products, this is not a minor issue: identity fidelity is a quantifiable metric (face-recognition embedding similarity, cross-age / cross-style stability), and the open-source ecosystem has long built evaluation baselines with SSIM / FID / KID / LPIPS and face-recognition accuracy (e.g. community measurements where PuLID face-recognition accuracy is 91%, InstantID 84%, IP-Adapter FaceID 76—82%; this set is low-confidence community measurement, for order-of-magnitude reference only). Miaoya neither publicized its own evaluation methodology nor put evaluation capability into users' hands, so users could not judge "whether the avatar generated this time looks more like me than last time" and could only vote with their feet.

5.7 L6 Governance and Safety Layer

Miaoya's record at the governance layer is negative first, then corrected:

• User-agreement controversy (2023): early on, its user-agreement terms sparked controversy — the focus was the boundary of the platform's rights to use users' uploaded photos and generated content. Officials later apologized and revised the agreement, promising that photos would be used only for this avatar creation, that original uploaded photos would be automatically deleted after completion, and that anonymization would be supported. The details of this incident come from third-party accounts and are marked [To be verified].
• Identity-marking compliance gap: no public explanation of any implementation of generated-content marking (explicit marking / implicit metadata marking / digital watermark) has been found for Miaoya, marked [To be filled].

In terms of compliance applicability, Miaoya's product form falls within the scope of multiple regulatory requirements:

• Measures for the Labeling of AI-Generated Synthetic Content (National Cyberspace Administration Communication Note 〔2025〕No. 2, effective 2025-09-01), Article 4: when service providers provide functions such as downloading, copying, and exporting of generated synthetic content, they shall ensure the file contains the required explicit labeling; Article 5: they shall add implicit labeling to the file metadata (including attribute information of the generated synthetic content, the service provider's name or code, content number, etc.), and are encouraged to add digital-watermark forms of implicit labeling. Miaoya's core paid point is precisely "HD download", which falls directly within the scope of Article 4.
• Civil Code of the People's Republic of China, Article 1018: a portrait is a "recognizable external image of a specific natural person reflected on a certain medium"; Article 1019: no organization or individual may infringe another's portrait right by defaming, defacing, or forging through information-technology means or otherwise, and without the portrait-right holder's consent, may not produce, use, or publish the portrait-right holder's portrait. Although the digital avatar is trained on the user's own portrait, once it is used to generate scene images different from the user and distributed externally, it enters the review scope of portrait right and the authorization chain.
• A Beijing Internet Court judgment effective 2026-03 established two key rules: identifiability is the core criterion for determining infringement (an AI face-swapped image need not be fully identical to the original portrait; if the general public can identify the person, it constitutes use of a specific natural person's portrait); burden-of-proof shift (a defendant claiming "accidental AI resemblance" must reproduce the creative process; failure to reproduce leads to the adverse consequence of failing to prove). The judgment also clarified that "technological neutrality" is not a ground for exemption.
• Industry warning: on 2026-04-28, Jimeng AI — also an object of this group's research — was investigated and dealt with by cyberspace authorities for failing to effectively implement the labeling requirements for AI-generated synthetic content. This shows that the regulatory focus is on the export and distribution stages, not the model capability itself.

Miaoya's lesson: governance is not a documentation exercise for the compliance department, but an engineering implementation of L6 — the authorization chain, identity labeling, audit trails, and reproducible creation records, all indispensable. The user-agreement controversy is merely the first visible face of the missing L6.

5.8 Maturity Assessment and Failure Attribution

Conclusion: Miaoya is the archetype of "strong model, weak Harness". Its generation capability was top-tier in 2023 (otherwise it would not have topped the App Store overall chart), but four of the six Harness layers (L2 / L3 / L4 / L5) were blank or failed, and L6 was even a negative asset at a critical time.

These five layers did not fail independently but amplified one another: L4 lock-in prevented users from accumulating assets → missing L2/L3 prevented users from creating new value → users left → missing L5 prevented the platform from noticing quality decline → the L6 incident accelerated trust erosion. No matter how strong the model, it cannot fill the structural gaps of the Harness.

6. Actual Cases

6.1 Scale Data at the 2023 Outbreak Moment

The available scale data are all point-in-time values from 2023 and must be labeled as such; they cannot be cited as the current state:

• Topped the App Store overall chart in 2023-08;
• DAU exceeded 600,000;
• At peak, 4,000+ people queued for portrait generation.

The meaning of this data is to depict the industry backdrop of "the first wave of the AI portrait track": it proves that there is real and strong C-end demand for "AI-generating one's own image". The demand was real, yet the product failed to keep users — which precisely shows that the demand side was not the bottleneck; the supply-side engineering capability (Harness) was.

6.2 Ctrip "AI Travel Photography"

According to a third-party source, Miaoya once opened an API to cooperate with Ctrip on an enterprise-level "AI Travel Photography" service. This information has low confidence; no official announcement or Ctrip-side product page has been found to corroborate it, marked [To be verified].

Even if true, from a Harness standpoint this cooperation would be hard to scale: Miaoya has no public API contract, concurrency quota, callback mechanism, or error-code specification; B-side integration could only rely on bespoke per-case work, with marginal cost not declining with customer count. This is fundamentally different from "platformized B-side capability".

6.3 User-Agreement Controversy Incident

In 2023, Miaoya sparked controversy over its user-agreement terms — the core of the dispute was the boundary of the platform's rights to use users' uploaded photos and generated content. Officials later apologized and revised the agreement, making three promises: photos would be used only for this avatar creation, original uploaded photos would be automatically deleted after completion, and anonymization would be supported.

This incident is the earliest "identity-asset governance incident" record in this group's documents; its engineering meaning is that when a product takes users' biometric features (faces) as a core input, the authorization chain and the data lifecycle must be first-class citizens of L6, not hidden inside a lengthy agreement text. Subsequent products of the same kind (such as Jimeng restricting real-person material use and introducing digital-human avatar authentication from 2026-02) are precisely a productized response to this lesson.

6.4 Lifecycle Review

Miaoya demonstrated that an AI portrait product defined by "single-point viral hit + no workflow depth + no asset portability" is hard to retain. Its lifecycle can be summarized in four phases:

1. Demand-validation phase (2023-07 to 2023-08): extremely low acquisition cost + extremely strong novelty → explosive growth;
2. Experience-overdraft phase (2023-08 to 2023-12): queuing, uncontrollable outputs, agreement controversy → reputation decline;
3. Capability-stagnation phase (all of 2024): competitors rapidly caught up on zero-shot identity injection and editing capabilities, while Miaoya saw no architecture-level upgrade;
4. Exit phase (2025): after the 2025-09 version added AI retouching and outfit-plan customization, the team disbanded at the end of 2025-09 and the product was kept at minimum-level operation.

Of these, the third phase is the most alarming: Miaoya was not defeated by a specific competitor but was left behind by a "technology paradigm shift". When zero-shot identity injection (InstantID / PuLID) compressed "20 photos + hours of queuing" into "1 photo + second-level output", Miaoya's cost structure instantly went from "an acceptable price" to "an unreasonable burden". And it lacked the ability to complete this migration, because its entire product architecture (and even its business model) was built on the technical premise of "training-based avatars".

7. Summary

7.1 Strengths

1. First attempt to make "identity" into a mass-market product: in 2023 it already let ordinary users experience an identity-asset form of "invest once, reuse many times", which was pioneering.
2. Generation quality was top-tier among its peers: being able to top the App Store overall chart and drive social virality shows the output quality passed the test of mass-market aesthetics.
3. Extremely low acquisition cost: mini-program-first + social sharing is a classic paradigm for C-end AI products to acquire users cheaply.
4. Validated a real demand: 600,000 DAU proves that "AI-generating one's own image" is a real need, not a fake one.

7.2 Limitations and Root Causes of Failure

1. The technical route was left behind by a paradigm shift: the training-based avatar completely lost its cost advantage after zero-shot identity injection matured, and the architecture could not migrate smoothly.
2. L4 asset locked one-way: the identity asset could not be migrated, combined, or versioned, so user investment became sunk cost.
3. Missing L3 orchestration collapsed the experience: hours of queuing, no batch processing, no intermediate-state intervention.
4. Blank L5 evaluation: no quantitative identity-fidelity methodology; quality could not be regressed or demonstrated.
5. L6 governance was negative early on: user-agreement controversy + blank identity-marking mechanism.
6. No depth to the business model: one-time ¥9.9 buyout, lacking subscription anchors and subsequent consumption scenarios.

7.3 Applicability Boundary

7.4 Engineering Lessons for Successors

1. Identity assets must be portable: make identity an asset users can export, migrate, and version, not the platform's private state.
2. The technical route must be able to migrate across paradigms: treat "training-based" as a swappable implementation detail, not a premise of the product architecture.
3. L3 orchestration is the floor of experience: queuing, batch processing, and intermediate-state intervention are not optimization items but usability items.
4. L5 must quantify identity fidelity: no metrics means no regression; no regression means no trust.
5. L6 governance upfront: the authorization chain, identity labeling, and audit trails should be done before the product launches, not patched after a controversy.
6. The compliance focus is on export and distribution: Article 4 of the Labeling Measures directly targets "download, copy, export", and Jimeng AI being investigated on 2026-04-28 is the proof.

7.5 Selection Recommendation

Miaoya Camera has no current-selection value; its value lies in serving as a reference coordinate for evaluating other platforms. When evaluating any outfit-changing, face-swapping, or portrait platform, the following four criteria distilled from the Miaoya case can be used directly:

Any product that fails all four criteria, no matter how strong its model, should be treated as a high-risk target.

Information-Gap Statement

1. Authoritative source for the team's disbandment: only a third-party tools website states it was "formally disbanded at the end of September 2025"; no corroboration from tech media or business / judicial information has been located, marked .
2. Current App version number: two conflicting figures exist — v4.6.14 (updated 2026-08-18) and v4.3.1 (updated 2026-04-08), marked [To be filled].
3. Current operating status: whether the product still provides services and its scope have not been reliably found, marked [To be filled].
4. Pricing data: ¥9.9 / ¥29.9 and 2 Diamonds/time are all third-party consistent statements at the 2023 point in time, not archived official pages, marked .
5. B-side cooperation (Ctrip "AI Travel Photography"): a single low-confidence source, no official announcement confirms it, marked .
6. The specific terms of the user-agreement controversy: only third-party accounts of the dispute's focus; no original agreement text has been found, marked [To be filled].
7. Identity-marking compliance implementation: no public explanation by Miaoya of explicit / implicit labeling or digital watermarks has been found, marked [To be filled].
8. Official merchant cases and quantitative effectiveness data: none found; honestly marked as "not found" rather than replaced by vague wording such as "widely used in the industry".
9. Technical details of the "Tixiang" model (parameter count, architecture, training data): no public technical report has been found, marked [To be filled].

8. References

1. Sparkx Collection · Miaoya Camera product page (includes launch time, team-disbandment information, and pricing figures; third-party source, medium confidence). https://www.sparkx.zone/tools/174
2. Full text of the Measures for the Labeling of AI-Generated Synthetic Content — Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security, National Radio and Television Administration, 2025-03-14. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
3. Interpretation of the Measures for the Labeling of AI-Generated Synthetic Content — Gov.cn / Xinhua News Agency, 2025-03-16. https://www.gov.cn/zhengce/202503/content_7014281.htm
4. "From September 1, AI-generated synthetic content must carry a label" — CCTV, 2025-03-15. https://big5.cctv.com/gate/big5/news.cctv.cn/2025/03/15/ARTI36OOL0hP5mpvU5cDgo4L250315.shtml
5. "Technology is not an infringement shield; court rules 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
6. "e-case | e-trial: A short-drama character's AI face-swap 'closely resembles' a famous actor — accidental resemblance or deliberate infringement?" — contributed by the Beijing Internet Court, The Paper. https://www.thepaper.cn/newsDetail_forward_32799628
7. Baidu Baike · Jimeng AI (including entries on being investigated on 2026-04-28 for failing to implement labeling rules, the digital-human avatar authentication mechanism, etc.; secondary source, recommended to cross-check against official announcements). https://baike.baidu.com/item/%E5%8D%B3%E6%A2%A6App/67386767
8. InstantID official project page — InstantX Team / Xiaohongshu / Peking University (technical reference baseline for zero-shot identity injection). https://instantid.github.io/
9. CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models — arXiv 2407.15886 (terminology and metric baseline for the outfit-changing direction). https://arxiv.org/pdf/2407.15886
10. InstantID vs PuLID vs FaceID Comparison — Apatero (third-party measured comparison, low—medium confidence, for order-of-magnitude reference only). https://apatero.com/blog/instantid-vs-pulid-vs-faceid-ultimate-face-swap-comparison-2025
11. Miaoya Camera official site — Weixu Network Technology (Shanghai) Co., Ltd. https://www.miaoya.cn/