Kling AI Image (Kuaishou)

1. Introduction

1.1. Platform Overview

Kling AI is developed by Kuaishou Technology. The image line and the video line share the same product entry point and the same Credits metering system. In the AI Harness six-layer capability model, Kling's unique value lies in the fact that the L4 Memory & State layer has been productized in the most explicit way: it abstracts "people / products / objects" into Element Creation as first-class assets, and directly specifies the number that can be held per membership tier (30 / 50 / 150 / 500) — a rare design in this group where the "asset layer is explicitly priced and explicitly limited."

At the same time, Kling O1 Image follows the "generation as conversation" route: it does not distinguish generation from editing — upload reference → generate → iterate with text corrections, forming a multi-turn conversational loop. This pushes it one step further than Midjourney on the L3 orchestration layer, though it still falls short of workflow-level artifacts.

1.2. Model Lineage

Official and channel statements for Kling O1 Image: supports 18+ task types, 4K output, and up to 10 reference images. These parameters come from a third-party aggregator marketing page, not official.

1.3. Pricing System

1.3.1. International-Site Membership (official page, high confidence)

Official conversion basis: 660 Credits ≈ 3,300 images or 33 720p videos. By reverse inference, one image is on the order of about 0.2 Credit.

Credits top-ups: 3,500 Credits / $50; 7,500 Credits / $100; 16,000 Credits / $200.

1.3.2. China-Site Membership (secondary source, medium confidence)

Source: a research report screenshot from Guosheng Securities. The China site and the international site align in their Credits / inspiration-point tier design (660 / 3,000 / 8,000 / 26,000), but the pricing systems are independent. China-site prices [To be verified — refer to the real-time price in the App].

Basis note (listed in parallel across the group): Section 1.3.2 records the China-site membership prices as Gold ¥58 / Platinum ¥234 / Diamond ¥586 / Black Gold ¥1,149 (inspiration points 660 / 3,000 / 8,000 / 26,000); another China-site basis is recorded in 03-市场研究/04-AI-漫剧组/02-kling-drama.md: Gold ¥19 / Platinum ¥58 / Diamond ¥198 (inspiration points 660 / 2,000 / 8,000). The international-site tier ceiling also has two figures: the Ultra in this section 1.3.1 is $127.99/month, while the Ultra renewal price in 02-kling-drama.md is $159.99/month. Both China-site figures come from relayed public channels and have not been directly confirmed against the official pricing page; the discrepancy may correspond to different time periods or different package systems. This section does not adopt either one; please refer to the real-time price on the official page when citing.

2. Glossary

2.1. General AI Image Terms

2.2. Kling-Specific Terms

2.3. General Terms for Virtual Try-On and Face Swap

Note: In version 1.5, Kling added the AI Model capability, and supports the reuse of product and person assets through "Element Creation", so it has practical touchpoints for virtual try-on; the platform does not offer a standalone face-swap product capability.

3. Feature Description

3.1. Generation and Editing Capabilities

3.2. Natural-Language Instruction Editing

Kling's editing paradigm is "use words instead of painting": the user describes "change the coat to burgundy", and the model locates the target region itself and completes the change. This forms a three-generation interaction difference versus Midjourney's Vary Region (requires first selecting a region) and Jimeng's intelligent canvas (requires first painting):

Engineering implication: the masking step is a frequent source of human error and workflow interruption. Removing the mask means a single generation can be programmatically and continuously corrected, a substantive step by the L3 orchestration layer toward "automation". However, Kling has not yet solidified this chain into an exportable workflow artifact.

3.3. Element Creation and AI Model

• Element Creation (Element): fixes people, products, and objects as reusable subjects; the tier determines the count cap (30 / 50 / 150 / 500).
• AI Model (added in version 1.5): combined with Element Creation, it can support the reuse of virtual models in e-commerce scenarios.
• Daily inspiration points: members receive daily inspiration points (specific value [To be filled]).

3.4. Commercial Licensing and Concurrency

• Content generated on the free tier (Basic) is not permitted for commercial use — this is an explicit clause on the official membership page and is one of the rare free-tier commercial-use restrictions in this group that are explicitly written down by the vendor; pay special attention during selection.
• Standard and above include commercial-use rights (stated explicitly on the official membership page).
• Concurrency: membership tiers offer "unlimited concurrent tasks" + a dedicated fast-generation channel.
• Watermark: paid tiers remove the brand watermark.

4. Platform Architecture

图 4-1｜可灵在 AI Harness 六层能力模型中的实现

数据来源：基于本文分析绘制的示意图。

4.1. Dual-Model Parallel

The two lines run in parallel, showing that Kuaishou has not pushed all capabilities onto a single model but instead divides them by "consistency / conversational editing" and "photorealistic rendering / text". This is conceptually similar to FLUX.2's "model selection as orchestration" (klein realtime / pro production / flex layout / max highest quality), but Kling has not elevated model selection into an orchestratable routing strategy — users can only switch manually.

4.2. Text Encoder Trade-offs

Kolors uses GLM as its text encoder, differing from the T5 commonly used by Imagen / SD3. Using a natively Chinese large language model as the encoder is the underlying reason for its strengthened Chinese-English comprehension.

This fact comes from a third-party technical write-up; the original paper has not been located.

4.3. Open Format

• Web (klingai.com) + App.
• Enterprise-side API (Kuaishou Open Platform).
• Integrated into third-party platforms (e.g., vivago.ai) and aggregator platforms; the aggregator ecosystem contributes meaningfully to its distribution.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Kling's L1 design is the one in this group with the greatest "context layering" awareness:

• Input layer: up to 10 reference images (channel source; recommended to recheck against official documentation). Officially it is recommended to extract features from 5–10 reference photos at different angles, different lighting, and different distances.
• Layering mechanism (Identity Map): divides features into two categories —
• Stable identity markers: persistent attributes such as bone structure, salient features, and proportions;
• Transient attributes: attributes that change with the scene, such as expression, outfit, lighting, and makeup.
• Anchoring logic: only the former is anchored during generation, while the latter is allowed to vary freely.

This is exactly the last item — priority ordering — in the L1 definition of "retrieval, compression, caching, priority ordering" from the parameter card. Kling turns "what must stay consistent, what may change" into an explicit product semantic, rather than letting users tune weights themselves (contrast Midjourney's single---ow numeric knob).

Areas to strengthen: no structured prompt contract (contrast FLUX.2's JSON fields), no external knowledge retrieval (contrast Seedream 5.0's web retrieval and Nano Banana's Search Grounding).

5.3. L2 Tools & Execution Layer

• Toolset: generation, editing, style transfer, super-resolution, video extension, Element Creation.
• Invocation form: uniformly invoked via natural language within the same session; tool boundaries are invisible to the user.
• No open tool registration: no Function Calling support, no MCP Server (contrast Runway, which released an MCP Server in 2025-06).
• Cost basis: Credits are a unified metering unit shared by images and video. The official conversion of 660 Credits ≈ 3,300 images gives decent cost estimability, but there is no pre-generation cost-estimation endpoint (contrast Leonardo.ai's Pricing Calculator).

5.4. L3 Orchestration & Control Layer

"Generation as conversation" is Kling's core contribution at L3:

Three steps are removed: re-uploading, masking/painting, and whole-image re-rolling. This brings two direct benefits:

1. Context is not lost: the reference images and the generated result both remain in the session, avoiding consistency breaks caused by "starting a new round".
2. Iteration can converge: each change touches only one spot, converging to the target over multiple rounds rather than rolling the dice again each round.

Still missing: no DAG, no sub-agents, no interruption recovery, no workflow export or version control. Compared with ComfyUI turning workflows into version-controllable JSON graphs and Meitu Design Studio using Agent Teams for multi-agent division of labor, Kling's L3 remains at "human-driven multi-turn conversation" and has not yet entered "system-driven automatic orchestration".

5.5. L4 Memory & State Layer

Element Creation is the most straightforward L4 productization design in this group:

Kling writes "asset count" directly into the membership entitlements table, meaning assets are product value. This makes commercial sense, but it also carries the same risk as Miaoya: assets are locked inside the platform. The difference is that Kling's asset form (person / product / object subjects) is more general than Miaoya's "digital avatar", and the platform remains in continuous operation.

Gaps: no Checkpoint, no artifact version management, no model-version locking endpoint, no documentation of an asset export mechanism.

5.6. L5 Evaluation & Observation Layer

• Kolors ranked 2nd globally in subjective overall score and 1st in subjective image quality on BAAI FlagEval. This result is a point-in-time leaderboard value; it must be timestamped and should not be treated as a continuously valid conclusion.
• No public official Eval Set, Golden Dataset, or regression set.
• Observation basis: Credits consumption. No pre-generation cost estimation, no A/B mechanism.

5.7. L6 Governance & Security Layer

• Tiered commercial-use rights: the free tier is not for commercial use (stated explicitly on the official membership page), and Standard and above include commercial-use rights. This is one of the clearest tiered commercial-use designs in this group.
• Watermark: paid tiers remove the brand watermark.
• Portrait and identity risks: the portrait-rights risks posed by real-person footage and identity consistency (Identity Map explicitly uses real people as reference subjects) are governed by the platform's membership agreement; no structured identity authorization or guardrail mechanism was found (contrast Nano Banana's public-figure portrait restrictions and SynthID's mandatory watermarking).
• Compliance with the Labeling Measures: no official public description of explicit / implicit labeling implementation was found. [To be filled]

5.8. Maturity Assessment

Kling is a "medium-strong model + medium-strong Harness" form, whose greatest features are the most straightforward L4 assetization (Element Creation limited-quantity pricing) and L3 generation-as-conversation (removing masking and re-rolling). It is closer to a deliverable form than Midjourney, but still has a clear gap from Meitu Design Studio's Agent Teams and ComfyUI's version-controllable workflows. If Miaoya is taken as a failed sample of "having assets but not transferable", then Kling is the control group of "having assets and an active platform" — whether assets are transferable and whether the platform stays operational together determine the real value of L4 investment.

6. Practical Cases

6.1. Official Customer-Case Search Results

This search did not find any brand customer case officially published by Kling with quantified performance data. Per this group's uniform basis, it is marked as [No results], without substituting vague phrasing such as "widely used in the industry" and without fabricating cases.

6.2. Channel-Partner Workflow Example (not cited as a case)

Workflow example provided by a third-party channel partner (vivago.ai): an AI virtual influencer — 5 reference images → 120 scenes/month, no photographer or model scheduling, with commercial-use rights included in the Plus plan.

This content is channel marketing material, not an official Kling release, with no third-party verification; this report does not cite it as a case, and treats it only as a "typical usage" reference. The quantified figure within it (120 scenes/month) should not be re-quoted as performance data.

6.3. Verifiable Usage and Allocation Data

Verifiable factual material (not performance data):

1. Official conversion basis: 660 Credits ≈ 3,300 images or 33 720p videos (official membership page, high confidence) — usable for per-unit cost estimation.
2. Element Creation tier limits: 30 / 50 / 150 / 500 (official page, high confidence) — usable for capacity planning.
3. Free tier not for commercial use (official page, high confidence) — usable for compliance determination.
4. Kolors open source (2024-05) — usable for local deployment and secondary-development evaluation.

7. Summary

7.1. Strengths

1. Identity Map's context-layering design: explicitly distinguishes stable identity markers from transient attributes, the clearest semantic design at the L1 layer in this group.
2. Generation as conversation: removes masking and re-rolling; iteration can converge and human error points are notably reduced.
3. Element Creation: assets become first-class and are limited-priced, making capacity planning predictable.
4. Clear tiered commercial-use rights: free tier not for commercial use is written down explicitly by the vendor, making compliance determination simple.
5. Officially given Credits conversion basis: 660 Credits ≈ 3,300 images, so per-unit cost is estimable.
6. Kolors open source: leaves a path for local deployment and secondary development.

7.2. Limitations and Applicability Boundaries

Applicability boundaries: suited to content production that needs "reusing the same person / the same product across multiple scenarios" (virtual influencers, e-commerce models, IP character extension); not suited to enterprise-grade pipelines requiring programmable orchestration, reproducible regression, and cross-system auditing.

7.3. Selection Recommendations

1. Do capacity planning before choosing a tier: back out the tier from two metrics — "monthly subject count + monthly image count". Subject count is a hard constraint (30 / 50 / 150 / 500), and image count can be topped up with single Credits purchases.
2. Must upgrade before commercial use: free-tier output is not for commercial use and must not be used in any external-facing material.
3. Build your own reproducible record: record the four-tuple of "reference-set hash + instruction sequence + model version + Element ID" as regression and evidence material.
4. Parameter promises must defer to the official source: parameters such as 10 reference images, 4K, and 18+ task types come from channel marketing pages; confirm with Kuaishou officially before signing.
5. For reference images of real people: must obtain written authorization from the rights holder and keep the authorization chain; platform agreements cannot replace rights-holder-to-rights-holder authorization.

7.4. Compliance Notes

• Articles 4 and 5 of the Measures for the Labeling of AI-Generated Synthetic Content (Office of the Central Cyberspace Affairs Commission [2025] No. 2): when service providers offer download, copy, and export functions, they shall ensure explicit labeling and shall add implicit labeling to the file metadata. No public implementation description was found on Kling's side; when used by enterprises, you should build the labeling-injection step yourself.
• Article 1019 of the Civil Code of the People's Republic of China: it is prohibited to infringe on another person's right to portrait by means such as forgery using information-technology methods. Identity Map builds stable identity markers from real-person reference photos, a highly sensitive form of identity reuse.
• A judgment effective in 2026-03 at the Beijing Internet Court established "identifiability + burden-of-proof transfer": the result need not be identical to the original portrait; as long as the public can identify it, it constitutes use of a specific natural person's portrait; a defendant claiming "coincidental resemblance" must reproduce the creation process. When anchoring identity with multiple reference images, once a generated result closely resembles a natural person, the burden of proof lies with the user.

Information Gap Statement

1. Original source of Kolors using the GLM text encoder: only mentioned in third-party technical write-ups; the paper has not been located.
2. Kling O1 Image's "18+ task types / 4K output / 10 reference images": from a third-party aggregator marketing page, not official.
3. China-site membership prices (¥58 / ¥234 / ¥586 / ¥1,149): from a broker research report screenshot, not the official page.
4. Daily inspiration-point amount: not found. [To be filled]
5. Official customer cases and quantified performance data: not found. [No results]
6. Details of explicit and implicit labeling implementation under China's Labeling Measures: no official public description found. [To be filled]
7. Seed locking, CFG exposure, model-version locking endpoint: no official description found. [To be filled]
8. Whether Kolors 2.1 is open source: no clear statement found.
9. Specific date of the FlagEval leaderboard: the leaderboard release date was not found; when citing, it must be labeled as a point-in-time value.
10. Pricing and concurrency specifications for the enterprise-side API: no official API pricing page found. [To be filled]

8. References

1. Kling AI Membership Plans (official page, including tiers, Credits conversion, and commercial-use terms) — Kuaishou, 2026. https://app.klingai.com/global/membership/membership-plan
2. Kling AI Official Website — Kuaishou. https://klingai.com/
3. Measures for the Labeling of AI-Generated Synthetic Content (Office of the Central Cyberspace Affairs Commission [2025] No. 2) — Cyberspace Administration of China, Ministry of Industry and Information Technology, Ministry of Public Security, National Radio and Television Administration; issued 2025-03-14, effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
4. Civil Code of the People's Republic of China, Articles 1018 and 1019 — National People's Congress, 2020. (Articles cited in the body text; no official link)
5. Economic Information Daily · "Technology Is Not a Shield against Infringement — How the Court Found AI 'Face Theft'" — Xinhua News Agency / Economic Information Daily, 2026-04-17. http://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm
6. InstantID official project page (zero-shot identity injection, as the comparison baseline for Identity Map) — InstantX Team / Xiaohongshu / Peking University. https://instantid.github.io/
7. DeepWiki · SDXL_EcomID_ComfyUI 5.4 Comparison (comparison of EcomID / InstantID / PuLID control mechanisms) — Alimama. https://deepwiki.com/alimama-creative/SDXL_EcomID_ComfyUI/5.4-comparison-with-other-methods
8. ComfyUI official workflow "Virtual Character Try-On - All-in-One" (comparison baseline for version-controllable workflows) — Comfy Org. https://comfy.org/zh/workflows/templates_rob_fashion_shoot_vton-4in1.app/
9. CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models — arXiv 2407.15886 (virtual try-on technical approach and evaluation metrics). https://arxiv.org/pdf/2407.15886
10. Black Forest Labs · FLUX.2 Overview (comparison baseline for model routing and fixed-snapshot endpoints) — BFL, 2026. https://docs.bfl.ai/flux_2/flux2_overview
11. Leonardo.Ai Pricing (official page; comparison baseline for the Pricing Calculator and capacity constraints) — Leonardo Interactive. https://www.leonardo.ai/pricing