AI Image Group Overview and Cross-Comparison

1. Group Overview

1.1. Research Question and Unified Framework

This group conducted market research on 16 platforms in the AI image domain (including 1 open-source ecosystem), using the AI Harness six-layer capability model defined in the project parameter sheet as the analytical framework:

The core research conclusion of this group can be summarized in one sentence: the competition in the AI image industry has shifted from "whether the model produces good-looking results" to "whether the generation process is predictable" — Harness capabilities such as reference-image slots, reproducible endpoints, workflow version control, and content labeling are becoming the real differentiators among platforms.

1.2. Data Timeline and Evidence Grading

• All platform data are snapshots as of the retrieval window of 2026-09-11 to 2026-09-12. The AI image domain iterates extremely quickly (within the retrieval window alone, Midjourney V8.2 launched, Qwen-Image 3.0 went closed-source, and the Kling image line entered O1 / 3.0 Omni), so version numbers, pricing, and ranking positions should be re-verified in real time before citing.
• Evidence grading follows the retrieval-report standard: official sources (official websites, official documentation, financial reports, court rulings) > authoritative media > third-party reviews/communities (low-to-medium confidence, all annotated in the main text).
• The "information gap statement" at the end of each document lists data that was not retrieved or lacks sufficient confidence; platforms with no public quantified case studies are truthfully marked as "not retrieved," never replaced by vague wording.

2. AI Image Market Panorama

Divided by product proposition and user segment, this group's 16 samples cover eight niches:

Three structural trends run across all tracks:

1. Reference images have become the main battleground: from Midjourney's --oref to Nano Banana Pro's 14 references + 5-person consistency, the ability to "put brand, people, and products into context" determines the ceiling of commercial value;
2. The orchestration layer is moving up fast: Meitu Agent Teams, LOVART design agents, and ComfyUI workflows are three parallel forms that collectively turn the "multi-step production chain" into a product;
3. The tension between open and closed source is becoming public: Qwen-Image 3.0 withdrawing its weights (2026-07) and FLUX.2 klein 4B sticking to Apache 2.0 form a contemporaneous contrast, making the sustainability of the open route a must-answer question in platform selection.

3. 16-Platform Cross-Comparison Matrix

注：L1~L6 为六层 Harness 成熟度的定性评级（弱 / 中 / 中强 / 强 / 最强），依据各篇正文的逐层分析汇总；LiblibAI、LOVART、WeShop 三行依据组内对应文档的六层小结；PicCopilot 行证据最弱（其官方能力未获专题检索支持，仅框架性判断成立）。

4. Overall Six-Layer Harness Maturity Assessment

图 4-1｜AI 图像平台的六层 Harness 能力模型

数据来源：基于本文分析绘制的示意图。

4.1. Two Industry-Wide Weaknesses: L4 and L5

Putting all 16 samples side by side, the clearest industry conclusion is: the industry as a whole is weak in L4 (reproducibility and asset persistence) and L5 (lack of standards for effect evaluation).

• The industry-wide L4 gap: most platforms do not treat "a user's generation history, parameters, reference images, and model versions" as portable first-class assets. Miaoya Camera is the extreme case — the digital avatar is locked inside a single product and cannot be migrated, which is considered one of the engineering root causes of its team disbanding in 2025-09. The three that do this best (Meitu's creative asset reuse, Runway Brand Kits, and ComfyUI workflows + LoRA all owned by the user) happen to be the samples with the healthiest commercialization or the most active ecosystems.
• The industry-wide L5 gap: almost no platform publishes an official Eval Set or regression set. Industry evaluation mainly relies on three substitutes: vendor self-assessed leaderboards (Kling FlagEval, Qwen-Image topping 9 benchmarks — and notably Qwen's own self-assessed leaderboard shows an unfavorable fifth place, a rare exception), reverse-engineering from business metrics (Meitu's compute-point consumption and ARPPU multiples), and subjective human evaluation (Midjourney). The e-commerce vertical is the only scenario that turns evaluation into a hard gate (three-tier SKU consistency inspection), which precisely shows: without external hard constraints, vendors have no incentive to build an evaluation layer.
• Reproducibility is closing the L4 gap: FLUX.2's fixed snapshot endpoints, ComfyUI's JSON workflow version control, and Midjourney's seed parameter system together turn "reproducing with the same parameters" from a wish into an engineering switch — this is the most valuable industry convergence this group observed.

4.2. Extreme Samples Across the Six Layers

4.3. Reference-Image Capability Comparison (L1 Core Metric)

4.4. Pricing Quick Reference

5. Platform Selection Decision Tree

A directly executable decision tree for each requirement scenario follows (for platforms at each branch endpoint, see the corresponding document):

你的核心场景是什么？
│
├── 电商换装 / 商拍
│   ├── 服饰类目、要固定模特与换装深度 → WeShop（15 篇）
│   ├── 跨境营销素材广度（多语言、模板、视频本地化）→ PicCopilot（16 篇）
│   ├── 批量设计工作流 + Agent 化 + 数据可核验 → 美图设计室（05 篇）
│   └── 成本敏感、有 GPU 运维能力、可自建合规 → 开源生态 CatVTON / Re-CatVTON（12 篇）
│
├── 品牌创意 / 营销视觉
│   ├── 美学优先、可接受弱编排 → Midjourney（01 篇）
│   ├── 需要联网事实与最强治理 → Nano Banana / Gemini（04 篇）
│   └── 中文语境、与短视频生态协同 → 即梦（02 篇）/ Seedream（10 篇）
│
├── 人像写真
│   ├── C 端单次写真 → 妙鸭相机（06 篇，注意其运营收缩状态）
│   ├── 需要主体/身份长期一致 → 可灵 O1 Identity Map（03 篇）/ Nano Banana 多参考（04 篇）
│   └── 有技术团队、要零样本身份注入 → 开源 InstantID / PuLID / EcomID（12 篇）
│
├── 换脸娱乐 / 影视后期
│   ├── 低分辨率快速迭代、自带 NSFW 检测 → ReActor（12 篇）
│   └── 高分辨率身份保持 → PuLID / EcomID（12 篇）；视频侧 → Runway（08 篇）
│
├── 开源自建 / 私有化
│   ├── 图内有字（中文排版、信息图）→ Qwen-Image 1.0/2.0 权重（11 篇）
│   ├── 排版文字 + 可复现生产 → FLUX.2 flex / 固定快照端点（09 篇）
│   ├── 换装批量管线 → CatVTON（2.26GB 显存）（12 篇）
│   └── 无 GPU、要社区模型与工作流 → LiblibAI 云端（13 篇）
│
└── 设计中台 / 成套资产交付
    ├── 品牌全案（Logo 到短视频一揽子）→ LOVART（14 篇；注意星流国内版 2026-10-10 停服）
    └── 生成 API 深度集成进自家产品 → Leonardo PAYG（07 篇）/ 火山引擎 / 百炼 API

Three universal selection checks (regardless of which branch): ① whether the reference-image limit and character-consistency mechanism meet the scale of your business assets; ② whether a reproducibility means is provided (seed / fixed endpoint / workflow file); ③ whether L5 (consistency evaluation) and L6 (labeling compliance) are shouldered by the platform or by you — this is the real price difference between commercial platforms and the open-source ecosystem.

6. Compliance Chapter

AI image compliance constraints in China have completed the closed loop of "legislation → regulatory enforcement → judicial adjudication," and all documents in this group share this chapter as their baseline.

6.1. Regulatory Framework: the Labeling Measures and Deep Synthesis Rules

• The Measures for Labeling AI-Generated Synthetic Content (Guoxinban Tongzi Issue No. 2 of 2025, jointly issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration, effective 2025-09-01):
• Article 4: when providing download, copy, and export functions for generated synthetic content, explicit labeling that meets the requirements must be ensured in the file (for images, adding a prominent prompt label at an appropriate position);
• Article 5: implicit labeling must be added to the file's metadata (generated-synthetic attributes, the service provider's name or code, content number, etc.), and adding digital watermarks is encouraged;
• Article 6: dissemination platforms must verify the implicit labels in the metadata and handle them in three tiers (labeled / user-declared / traces detected), and append dissemination-attribute information;
• Article 10 (red line): no organization or individual may maliciously delete, tamper with, forge, or conceal labels, nor provide tools or services for others to do so.
• The supporting mandatory national standard, Cybersecurity Technology: Methods for Labeling AI-Generated Synthetic Content, was released at the same time.
• The Provisions on the Administration of Deep Synthesis Services in Internet Information Services: Article 16 (implicit labeling) and Article 17 (deep synthesis services such as face replacement and face generation that could cause public confusion or misidentification must carry prominent labeling).
• The Interim Measures for the Administration of Generative AI Services, Article 12: requires labeling of generated content.
• Current compliance implementation in this group: apart from Nano Banana (SynthID invisible watermark + C2PA + tiered visible watermarks), no other platform has publicly disclosed labeling implementation; the open-source ecosystem (Doc 12) entirely requires users to build their own labeling components. This is a must-check item in platform selection.

6.2. Civil Boundaries: Portrait Rights Clauses of the Civil Code

• Article 1018: a portrait is "the recognizable external image of a specific natural person reflected on a certain medium" — full identity is not required; "recognizable" alone qualifies for protection;
• Article 1019: no organization or individual may infringe another's portrait right through defacement, smearing, or forgery using information-technology means; no portrait may be produced, used, or published without consent — AI face swapping directly falls under "forgery using information-technology means";
• Article 1020: the fair-use circumstances (personal study, news reporting, official duty, public environment, etc.) do not include commercial face swapping;
• Article 1023: a natural person's voice is protected by reference to the portrait-right rules (voice cloning / speech synthesis are likewise constrained).

6.3. Judicial Implementation: Beijing Internet Court's March 2026 Ruling

In a case where a well-known actor sued a short-drama producer and broadcaster over AI face-swap portrait-right infringement (effective 2026-03), four adjudication points have become the industry compliance baseline:

1. "Recognizability" is the core adjudication standard: an AI face-swap image need not be completely identical to the original portrait — as long as the facial contours and features are highly similar and the general public can recognize them, using a specific natural person's portrait is established;
2. Shifting of the burden of proof: a defendant claiming "coincidental AI resemblance" must reproduce the creation process; failure to reproduce incurs the adverse consequence of failing to prove — especially unfavorable to enterprises using open-source workflows (many parameters, hard-to-reproduce process), which should proactively retain generation logs and parameter snapshots;
3. Copyright authorization cannot absorb portrait rights: obtaining authorization to the right of communication through information networks is not equivalent to authorization to use a portrait; a broadcaster that fails to fulfill reasonable review obligations likewise commits infringement;
4. "Technological neutrality" is not a ground for exemption, and an extremely short duration of the disputed clip does not constitute a defense either.

6.4. Regulatory Enforcement: Jimeng AI Investigated and CCTV Naming

• Jimeng AI (2026-04-28): was investigated and penalized by the cyberspace authorities for failing to effectively implement the requirements for labeling AI-generated synthetic content (per the Baidu Baike entry; re-verify against the official announcement). This is the only head-platform enforcement case within this group's research window, proving that enforcement of the Labeling Measures has reached first-tier big-vendor products.
• LiblibAI (2026-04): was named by CCTV for serious regulatory loopholes in several AI software products; "Libu Libu AI" was said to be able to generate improper content using veiled prompts while remaining normally downloadable from app stores (re-verify against the regulatory announcement).
• The two cases jointly show: the L6 governance layer is no longer a work item that can be retrofitted later, but a precondition for a platform's survival.

6.5. Criminal Risk Samples

• Mr. Wang of Zhangjiagang, Jiangsu: AI face-swap defamation case — convicted of defamation, one year and three months in prison;
• Mr. Sun of Ningyang, Shandong: AI face-swap combined with voice synthesis in an online-dating fraud case — convicted of fraud, three years' imprisonment with a four-year suspended sentence.

6.6. Compliance Point Comparison Across This Group's 16 Platforms

6.7. Compliance Self-Check Checklist

Any team that uses AI image capabilities to deliver content externally should verify each item before launch:

1. Has the output been given explicit labeling (prominent in-image notice)?
2. Does the file metadata contain implicit labeling (attributes, service-provider code, content number)? If the platform does not provide it, have you built and injected it yourself?
3. If real-person imagery is involved: has portrait authorization been obtained? Does the generated likeness pass the "recognizability" screening (err on the side of caution)?
4. For face-swap / face-generation uses, is prominent labeling applied per Article 17 of the Deep Synthesis Rules?
5. Have you retained a complete record of the reproducible creation process (to address the burden-of-proof shift)?
6. For e-commerce scenarios: do generated images and the physical SKU pass the "structure–attribute–compliance" three-tier consistency inspection (to prevent false advertising)?
7. Is there any misuse of capabilities such as label removal or watermark removal that touches the Article 10 red line of the Labeling Measures?

7. Document Navigation Index

This group contains 17 documents in total (16 platform studies + 1 group overview):

7.1. Methodological Boundaries and Usage Notes

To help those citing this group's conclusions use them correctly, the following methodological boundaries are explained:

1. Ratings are qualitative judgments, not quantitative scores: the L1–L6 ratings (Weak / Medium / Medium-strong / Strong / Strongest) in the Section 3 matrix are aggregated from each document's layered analysis, with the basis given in the corresponding document's 5.x sections; a subjective elasticity of ±1 level is allowed when comparing across platforms, and it is not recommended to convert the matrix directly into numeric scoring.
2. Evidence strength is asymmetric: among the 16 samples, Meitu (financial reports), Google (official pricing and cases), and Volcano Engine / Alibaba Cloud / Runway / Leonardo / FLUX (official documentation) have relatively strong evidence; some key data of PicCopilot, LOVART, and WeShop is missing (listed in each document's information-gap statement). Official channels must be re-verified before citing the PicCopilot row's conclusions.
3. Point-in-time sensitivity: version numbers and pricing have extremely short half-lives. Within the retrieval window alone, this group observed three major changes (Midjourney V8.2, Qwen-Image 3.0 going closed-source, and the Kling image line entering O1); citations should note the "2026-09 snapshot."
4. Trade-offs from the Harness perspective: this group deliberately downgraded "model quality" (aesthetics, image fidelity) to a background variable, focusing on the engineering carrier layer. If a reader's main decision variable is output aesthetics, they should additionally consult third-party blind-test leaderboards beyond these documents; this group does not provide conclusions in that dimension.
5. Case scope: apart from Meitu (financial-report data) and Google (THG Ingenuity / Max Fashion / L'agence), no official customer cases with quantified effectiveness data were retrieved for the other platforms; each document's main text marks this truthfully, and citations must not substitute vague wording such as "widely used in the industry."