Midjourney（V8.2）

1. Introduction

1.1. Platform Overview

Midjourney is developed by Midjourney Inc. (USA), which positions itself as a "community-funded research lab of about 60 people." In the AI Harness six-layer capability model, Midjourney is a typical sample of extremely strong model capability, extremely thin engineering layer: it invests almost all of its R&D resources in "image quality and aesthetics," while leaving engineering capabilities such as orchestration, state, evaluation, and governance for users to handle manually.

This trade-off is clearly visible in this research: the official documentation homepage announces the current version as "V8.2 is Live!" and positions it as an "update of aesthetics, image quality, and Personalization," with officials describing V8.2's images as "more creative, bolder, more mature, and more cutting-edge." By comparison, at the same time, Black Forest Labs' FLUX.2 already provides an explicit distinction between preview and fixed-snapshot endpoint types in its official documentation, to meet reproducibility and compliance-model-stability needs. The difference in investment in "engineering predictability" between the two is the core observation thread of this report.

1.2. Version History and Open Form

Midjourney's distribution form underwent the most significant engineering change in this research cycle for this group: the primary platform migrated from Discord to the Web editor. The old mental model of "Midjourney = a Discord bot" is no longer valid and should not be relied upon when writing or selecting technology.

• Discord era: driven by chat commands (/imagine), all generation results accumulated in the channel feed, with very weak search and asset-management capabilities.
• Web editor era: the web end hosts generation, editing, upscaling, asset library, folders, and team comments, while Discord is downgraded to an optional entry point.
• No mobile: no native app is supported, only the browser can be used, which imposes a hard constraint on usability in mobile-office scenarios.

The version timeline should be used with caution. The official documentation only confirms "V8.2 is Live," while the specific release dates of V7 / V8.1 / V8.2 come from a single secondary website (pixmind.io) citing docs.midjourney.com. This report presents the timeline with a layered approach — "officially confirmed version exists, third party provides dates" — and the date portion is marked [To be verified].

1.3. Pricing and Commercial Licensing

• No permanent free tier: the Free tier does not exist (consistent across multiple sources).
• Commercial licensing: paid tiers include commercial rights; third-party sources claim businesses with annual revenue above $1 million must select a higher tier. This claim [To be verified, subject to the official Terms of Service].
• Third-party relay APIs: on the market there are relay services quoting $0.04~$0.159 per image (varying with Draft / Fast / Turbo tiers). The source is a third-party website, of low confidence, and is not cited as fact — it only illustrates the ecosystem's current state.
• Inference tiers: Draft Mode (about 10x speed, about half the GPU cost), Fast Mode, Turbo, and Relax Mode (unlimited but low-priority queue, available on Standard tier and above).

2. Glossary

2.1. Common AI Image Terms

2.2. Midjourney-Specific Terms

2.3. Common Try-On and Face-Swap Terms

Note: Midjourney does not provide native virtual try-on or face-swap product capabilities. The following terms are listed with the group's unified cross-platform terminology for horizontal comparison; the mechanism closest on Midjourney's side is the --oref clothing and character feature transfer.

3. Feature Description

3.1. Generation and Editing Capability Matrix

3.2. Reference Image and Character Consistency Mechanisms

Midjourney's consistency approach is "two reference channels + one implicit profile":

1. --oref content channel: transfers the identity features of people, objects, vehicles, and clothing; the weight is controlled by --ow (1~1000, default 100, officials recommend below 400 under ordinary stylization conditions).
2. --sref style channel: locks in the reference image's aesthetic style.
3. Personalization implicit profile: aesthetic preferences accumulated from the user's historical scoring, enabled by default since V7 and substantially optimized in V8.2; it affects every generation without explicit parameters.

Engineering implication: the third channel is "invisible context." The same prompt yields systematically different results across different accounts, which is a negative factor for team collaboration and reproducibility — and Midjourney provides no public mechanism documentation for turning off or exporting this profile.

3.3. Parameter System

3.4. Known Engineering Constraints

The following constraints come from the official Omni Reference documentation and are hard limits that must be written into any technical solution:

1. Doubled GPU cost: generations using --oref cost 2x the GPU cost of standard generations.
2. Mode exclusivity: --oref does not support Draft Mode, Fast Mode, or Conversational Mode. This means "quickly testing directions" and "maintaining consistency" cannot both be achieved on Midjourney.
3. Parameters must be stripped before editing: for images generated with --oref, reference parameters must be stripped first before performing Vary Region, Pan, or Zoom Out, otherwise editing behaves abnormally.
4. Filtered results are still billed: content moderation filters some generations, and paid plans bill per request, so filtered requests may still consume quota.

4. Platform Architecture

图 4-1｜Midjourney 平台架构：分发入口、模型推理与接口集成三层

数据来源：基于本文分析绘制的示意图。

4.1. Model and Inference

• Model foundation: proprietary closed-source diffusion architecture; parameter count and technical report not disclosed, and no model card is provided.
• Inference acceleration: Draft Mode (distilled or low-step tier) + Fast / Turbo tiers + Relax Mode queue scheduling, forming a three-dimensional "speed—cost—priority" tier system.
• Capability-architecture mismatch: Midjourney has long been in the first tier for image generation quality, but at the architecture level it exposes almost no programmable contracts to the outside (no structured prompts, no model routing, no snapshot endpoints).

4.2. Distribution-Form Migration

This migration only solved "asset visibility" and did not solve "orchestration programmability" — what the web end provides is folders and comments, not workflows, DAGs, or sub-agents.

4.3. Interface and Integration Form

• Official public API: no official API documentation found.
• Ecosystem state: third-party relay services offer calls at $0.04~$0.159 per image (varying by Draft / Fast / Turbo tier). Such relays are not official contracts and cannot be relied upon as production dependencies for compliance or stability.
• Integration capabilities: no public documentation of an MCP Server, Webhook, or official SDK; integration with external environments such as Adobe and ComfyUI shows no official support.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Midjourney's context engineering is the layer most worth studying on this platform, characterized by "parameters are context":

• Content context: --oref carries identity, object, and clothing features, with weight --ow.
• Style context: --sref carries aesthetic style.
• Image context weight: --iw controls the weight of the input image in image-to-image.
• Implicit context: the Personalization aesthetic profile, injected long-term and invisible.

From the Harness perspective, this is a context system of strong expressiveness, weak controllability: users can connect visual context at very low cost, but cannot review, export, or freeze that implicit profile. In contrast to FLUX.2, which defines a structured prompt contract via official JSON fields (subject / background / lighting / style / camera_angle / composition), Midjourney's context organization relies entirely on the unstructured form of "prompt string concatenation" and cannot be programmatically validated or regressed.

5.3. L2 Tools and Execution Layer

• The tool set is fixed as "generation + editing + upscaling + variation + redraw + outpainting," invoked via CLI-style parameters.
• No tool-registration mechanism: users or third parties cannot register new tools, nor combine Midjourney capabilities into a larger tool chain.
• No MCP / Function Calling: Midjourney cannot be directly registered as a tool by external Agent environments (compare Runway, which released an MCP Server in 2025-06 and can be called directly by environments such as Claude).
• No sandbox concept: there is no mechanism for "post-processing generation results in a controlled environment."

Conclusion: L2 is a closed, limited tool set that only addresses "what can be done," not "how it is orchestrated."

5.4. L3 Orchestration and Control Layer

This is Midjourney's weakest layer:

• No workflows, no DAGs, no state machines, no sub-agent dispatch.
• No interruption and recovery mechanism: a generation cannot be interrupted; a failure can only be re-run.
• No batch processing: large-scale image output relies on manual repeated submissions or third-party scripts simulating front-end behavior.
• What the web end provides is folders and team comments — lightweight collaboration-layer capabilities, not orchestration-layer capabilities.

Engineering consequence: when a business needs tasks like "100 SKUs × 5 scene images, keeping the same model and the same lighting position," Midjourney can only be executed manually item by item, with no way to solidify the process into a reusable artifact. This is one reason this group treats L3 as the Harness watershed.

5.5. L4 Memory and State Layer

Reproducibility is recognized by this group as a first-class engineering property — FLUX.2 provides both preview and fixed-snapshot endpoint types, and ComfyUI turns workflows into version-controllable JSON graphs. On this axis, Midjourney only has a scattered --seed: the seed fixes only the noise, not the model version. After the model moves from V8.1 to V8.2, output with the same seed and the same parameter string will necessarily drift, and the platform provides no endpoint locked by model version.

5.6. L5 Evaluation and Observability Layer

• Officially, no Eval Set, Golden Dataset, or regression-set mechanism is visible.
• Quality judgment relies on human-eye subjective evaluation and community feedback.
• Third-party claims state that V7 reduces failed generations by 30%~40% compared to V6, but that data comes from an evaluation site, is not official, and is not recommended as a factual citation.
• No cost-observability measurement (compare Leonardo.ai's Pricing Calculator endpoint and Runway's unified $0.01/credit metering). Users cannot estimate cost before generating.

5.7. L6 Governance and Security Layer

• Content moderation: filters some generations; filtered results may still consume quota.
• Privacy: Stealth Mode is exclusive to the Mega tier.
• Commercial rights: graded by subscription tier; the annual-revenue-scale threshold claim is [To be verified].
• Identity and likeness guardrails: governance of face-swap and real-person materials is primarily community guidelines + moderation; no structured identity guardrails are visible (compare Nano Banana's SynthID + C2PA + restrictions on public-figure likenesses).
• China "Labeling Measures" compliance: no official public product-line documentation for "AI-generated labeling / watermarking" is visible. Whether it meets the explicit + implicit dual-labeling requirement is .

5.8. Maturity Assessment

Midjourney falls into the "strong model + weak Harness" form: L1 (reference context) and L4 (aesthetic profile) are done deeply, while L2, L3, and L5 are not engineered, and L6 is implemented through policy rather than mechanisms. It is one of the samples in this group's platforms with the largest gap between model capability and engineering layer, and the best entry point for understanding "why a strong model does not equal deliverability."

6. Real-World Cases

6.1. Official Customer Case Search Results

This search found no customer cases officially published by Midjourney with performance data. Midjourney's website only presents project themes (imagination / coordination / reflection / beauty / human flourishing) and does not list a customer roster.

Per this group's unified standard, we truthfully mark this as "no results" here, do not substitute vague phrasing such as "widely used in the industry", and do not fabricate cases.

6.2. Verifiable Use-Case Descriptions

Widely existing use-case descriptions (concept art, advertising creative, banners, product concept images) come from third-party evaluation sites, without performance data, and are therefore not cited as cases — they are referenced only as use illustrations.

6.3. Observations Valuable to the Harness Argument

Although it lacks commercial cases, Midjourney provides two engineering facts this group can directly cite:

1. Omni Reference's three hard constraints (2x GPU cost, mutual exclusivity with Draft/Fast/Conversational modes, and the need to strip reference parameters before editing) show that consistency capability on Midjourney is an "added cost item" rather than a "built-in default."
2. Personalization's implicit injection shows that, in the absence of an explicit context switch, "the same prompt yields different results in different people's hands" can become a systematic obstacle to team collaboration.

7. Summary

7.1. Strengths

1. Output aesthetics and image quality have long been in the first tier, and V8.2's official positioning clearly points to aesthetics, image quality, and Personalization.
2. Strongly expressive reference mechanisms: --oref / --sref can be stacked, enabling character and style transfer at a very low threshold.
3. Mature parameter system: --ar / --stylize / --chaos / --quality / --style raw and others give creators fine-grained control.
4. The web editor has completed the primary-platform migration, and the asset library and folders solve the asset-visibility problem of the Discord era.

7.2. Limitations and Applicability Boundary

Applicability boundary: suitable for exploratory creation whose deliverable is a "single polished image" (concept design, creative proposals, visual inspiration); not suitable for production pipelines requiring batch, reproducibility, auditability, and programmability.

7.3. Selection Recommendations

1. Creative exploration stage: Midjourney is the preferred choice, getting high-quality direction drafts at the lowest cost.
2. Production delivery stage: after finalizing in Midjourney, migrate the style and structure to a hosting layer with orchestration and snapshot capabilities (such as a ComfyUI JSON-graph workflow or FLUX.2's fixed-snapshot endpoints) to gain reproducibility and version control.
3. Before enterprise procurement: two things must first be confirmed — whether an official API exists, and the specific terms of commercial rights and the annual-revenue threshold in the official Terms of Service. Both are [To be verified].
4. Distribution within China: the implementation plan for the explicit + implicit labeling requirements under the "Labeling Measures" must be self-assessed; no public implementation documentation is visible on the platform side.

7.4. Compliance Notes

• Article 4 of the "Measures for Labeling AI-Generated Synthetic Content" (Guoxinban Tongzi [2025] No. 2) requires: when providing download, copy, and export functions for generated synthetic content, one shall ensure the file contains explicit labeling that meets the requirements; Article 5 requires that implicit labeling shall be added to the file metadata.
• Article 1019 of the Civil Code of the People's Republic of China: no organization or individual may infringe on another person's right of portrait by defacing, soiling, or using information technology to forge, or otherwise.
• The effective judgment of the Beijing Internet Court (2026-03) established the "identifiability + burden-of-proof shift" rule: a defendant claiming "the AI coincidentally resembles someone's face" must reproduce the creation process; if it cannot be reproduced, it bears the adverse consequence of failing to prove its claim. This rule is especially unfavorable to Midjourney usage that has "no workflow and no version records" — once litigation arises, users will find it hard to prove the generation process.

Information Gap Statement

1. Whether an official public API exists and version dates: no official API documentation found; all market quotes come from third-party relay sites ($0.04~$0.159 per image, low confidence).
2. Specific dates of V7 / V8.1 / V8.2: from a single secondary site (pixmind.io) citing docs.midjourney.com; officials only confirm "V8.2 is Live."
3. Commercial licensing and the annual-revenue threshold: the $1M threshold is a third-party claim. [To be verified, subject to the official Terms of Service]
4. Niji model-line version number: no reliable version information found. [To be filled]
5. Parameter count, technical report, model card: not disclosed by Midjourney. [To be filled]
6. Implementation details of explicit and implicit labeling under China's "Labeling Measures": no official public documentation seen. [To be filled]
7. Official customer cases and quantitative performance data: not found. [No results, must not fabricate]
8. Specific quotas for Relax / Fast / Turbo subscription tiers: no official quota table found. [To be filled]

8. References

1. Midjourney official documentation homepage — Midjourney Inc., 2026 (original: "V8.2 is Live!"). https://docs.midjourney.com/hc/en-us
2. PixMind · Midjourney V7 Review in 2026 (version timeline, secondary source) — 2026. https://www.pixmind.io/posts/midjourney-v7-review-and-guide
3. ToolCrush · Midjourney Review 2026 (open form and tiers, third party) — 2026. https://toolcrush.io/tool/midjourney
4. AI Tool Radar · Midjourney Guide 2026 (web migration and free tier, third party) — 2026. https://aitoolradar.io/guides/midjourney
5. "Measures for Labeling 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, issued 2025-03-14, effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
6. Interpretation of the "Measures for Labeling AI-Generated Synthetic Content" — Gov.cn / Xinhua News Agency, 2025-03-16. https://www.gov.cn/zhengce/202503/content_7014281.htm
7. Articles 1018 and 1019 of the Civil Code of the People's Republic of China — National People's Congress, 2020. (Provisions cited in the text, no official link.)
8. Articles 16 and 17 of the "Provisions on the Administration of Deep Synthesis of Internet Information Services" — Cyberspace Administration of China et al., 2022. (Provisions cited in the text, no official link.)
9. Economic Information Daily · "Technology Is Not an Infringement 'Shield'; Court Rules This Way on AI 'Face Theft'" (Beijing Internet Court effective judgment of 2026-03) — Xinhua News Agency, Economic Information Daily, 2026-04-17. http://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm
10. Black Forest Labs · FLUX.2 Overview (fixed-snapshot endpoints and structured prompts, as L4/L5 comparison baseline, official documentation) — BFL, 2026. https://docs.bfl.ai/flux_2/flux2_overview
11. ComfyUI official workflow "Virtual Character Try-On - All-in-One" (version-controllable JSON graph, as L3/L4 comparison baseline) — Comfy Org. https://comfy.org/zh/workflows/templates_rob_fashion_shoot_vton-4in1.app/
12. InstantID official project page (zero-shot identity injection, as L1 comparison baseline) — InstantX Team / Xiaohongshu / Peking University. https://instantid.github.io/