Hailuo, Kling, Runway research notes: connect model capabilities to the production process

"Latest" involves confirmation date, not model name

This note deals with official public materials obtained on 2026-10-04. We do not conclude that features announced in 2025 will be available in all regions and on all plans in 2026. Video generation services will update product UI, API, pricing, model name, and policies separately. Therefore, the term "latest model" must be read in conjunction with the announcement date, acquisition date, and confirmed aspects. This article is not an execution test or contract audit.

Read Hailuo's changes in 2025 in chronological order

At S2V-01 January 2025, MiniMax announced Subject Reference, which aims to identify subject identity in dynamic images from a single person reference. The provider itself notes that environmental deformation and weak prompt following can occur. This reservation is important. The claim that people can be maintained does not mean that places, objects, multiple people, or long-distance continuity can be automatically maintained. It is safer to consider only the reference person's face as a fixed asset, and protect clothing, props, and backgrounds with separate references and cut designs.

In Hailuo 02 June 2025, the company explained the efficiency improvement called NCR (Noise-aware Compute Redistribution), provided versions of 768p 6/10 seconds and 1080p 6 seconds, and introduced short format generation from text and images. The official ``2.5x training/inference efficiency'' is a comparison made by the company, not an independent comparison with any GPU or other companies' models. What we can definitely read from the production side is that the idea of ​​creating multiple 6-10 second clips and putting them together through editing also appears in the examples in the presentation. Rather than asking for 30 seconds at a time, the design of each cut is consistent with the length of the product.

1. 1Person Reference
2. 2S2V-01 Identity Test
3. 3Grade Face, Clothes and Background Separately

1. 1Hailuo02 Short Scale
2. 2Suggested cuts for 6-10 seconds
3. 3Reference the last frame to the next cut

1. 1Edit
2. 230-second continuity/sound/provenance

In August 2025, Start & End Frames and End Frame Only were announced, and the company explained that it supports start and end frames at 768p and 1080p, and start frames at 512p. What this solves is a problem in which the ``image you want to reach'' cannot be given as a condition. What it does not solve is the problem of making physics, fingers, objects, and people look natural every time. When you have a starting point and an ending point, don't try to cram in a difficult act with only one in between. Separate standing, walking, passing things, and turning around, and make the exit of each cut the entrance to the next.

Hailuo 2.3 announcement in October 2025 stated that it has improved complex human movements, stylization, microexpressions, and motion instructions for objects, and the provider claims that the Fast version reduces batch production costs by up to 50%. Rollouts to web, mobile, and Open Platform APIs are also included in the announcement. The production decision that should be made from this point is not to automatically select the Fast version as the final master, but to assign it to the rough stage for comparing composition, movement, and rhythm, and re-evaluate only the selected cuts with quality priority settings.

Hailuo Media Agent does not erase the process

Hailuo announced Video Agent as a beta in June 2025, and explained its evolution to Media Agent in October 2025. A batch workflow that automatically selects everything from ideas to materials, storyboards, editing, and audio can be effective for quickly seeing the first draft. However, whether the cut structure chosen by the model suits the brand, person's consent, product facts, and length of the publication site is a different judgment. Instead of viewing agent output as a finished product, pull out the shot list, reference history, audio sources, and text layers for human approval.

Before putting Runway and Kling in the same comparison chart

Runway's Gen-4 Research Announcement indicates the direction of generation using reference as the March 31, 2025 announcement. Although the public URL was confirmed at the time of acquisition, this note does not independently verify the full text of the page to determine performance numbers, current plans, and API availability. Regarding Kling, the text could not be confirmed due to robot restrictions during an automatic search when the official domain was acquired. Therefore, model numbers, performance rankings, fees, and licenses are not written in this book. Part of accurate technical research is not to fill in the gaps that cannot be investigated with word of mouth or aggregation sites.

If you want to make a comparison, place the same exam before lining up the service names next to each other. A 4-second shift with a person reference, a 4-second lift of the product, a 6-second connection between the start and end points, a background that moves horizontally, a scene that requires sound, and a reconfiguration into portrait orientation. Leave input asset rights, model/version, settings, output, failure type, and retry cost for each test. The physics,'' cinematic,'' and ``coherence'' of advertising can only be compared by breaking it down into individual points: hands, touches, backgrounds, people, and cut connections.

1. 1Same Reference Pack
2. 24 Fixed Tests
3. 3Scoring by Cut

1. 1Quality Assessment
2. 2Separate Identity/Contact/Camera/Background/Sound

1. 1Operational Evaluation
2. 2Duration/Retry/Price/Terms/Reproduction Record

1. 1Selection
2. 2Determined by total cost up to and including editing process

Check list before implementation

Before actually using Hailuo, Runway, or Kling, check whether the model is UI or API, whether it can be used in the region and account, handling of input images, people, and voices, storage and deletion, unit of charge, watermarking and provenance of output, and current terms of use. If you want to include a human face or customer material, check rights and consent first. What ultimately creates value is not applying a single model. This is the process of searching in short form, fixing it with references, completing the intention with editing, and leaving a history of changes.

Correct the dated status of this note

The Hailuo/Kling/Runway entries are historical research notes. A source date outside the note's observation boundary must never be treated as a result checked on 2026-10-04. In particular, keep later announcements out of a dated comparison until their publication date and original page are verified. In the 2026-09-04 to 2026-10-04 boundary, no primary Hailuo, Kling, or Runway release directly relevant to the cited controls was confirmed in this pass.

Use community reports only to find production questions—such as whether a single generation or an edited sequence is easier to control—and validate the answer with a current product contract and a small paid or local test.

Host availability and provider history must remain separate

Runway's current undated model guide can be consulted on 2026-10-04 to see its own selectable model surface. A host table that includes MiniMax H3 Max, Wan 3.0, Seedance 2.5, Veo 3.1, or Gen-4.5 does not establish a Hailuo or Kling provider launch, a direct API contract, or access in your account. The August 5 Runway notice about deprecated Gen-3/Veo 3 standalone tools likewise describes only Runway's surface.

Use a three-row evidence card for every candidate: original provider announcement and date; host product availability and observed date; your account test with input, output and export. Compare one bounded reference-to-shot task, retain the output and failure mode, and do not map a third-party name to a provider capability claim. This corrects the old note without rewriting its historical Hailuo and Kling sources.