Creatify Post-Trained MiniMax H3 for Ads and Cut the Cost to Four Cents a Second

--- title: Creatify Post-Trained MiniMax H3 for Ads and Cut the Cost to Four Cents a Second meta_title: Boreal-H3 Shows What Post-Training an Ad Model Buys meta_description: Creatify Labs post-trained MiniMax H3 into Boreal-H3, claiming higher ad pass rates and a $0.04 per second price roughly twelve times below Seedance 2.5. ---

The video model race has mostly been about which lab has the best general-purpose generator. Creatify Labs took a different route: start from someone else's strong base model, then train it specifically on the work an advertisement has to do.

Boreal-H3 is a MiniMax H3 model post-trained in partnership with MiniMax itself, launched October 2. It is the second entry in Creatify's Boreal family, following an original Boreal built for realtime generation at a penny per second.

What Boreal-H3 was trained to do

Creatify describes the training target as the three jobs an ad has to complete: keep the product from the reference image intact, deliver a natural creator performance, and follow the brief as written.

That framing is more specific than it sounds. A general video model optimizes for looking good. An ad model has to keep a label readable, preserve a package's shape across a clip, and make sure the person on camera stays recognizably the same person from the first frame to the last. Those are consistency requirements, not aesthetic ones, and they are exactly where general models tend to slip.

The company paired the launch with an Ad Agent powered by Claude Opus 5.5, which handles the creative direction side of the workflow.

The numbers and who produced them

Creatify reports Boreal-H3 at 133.3 on its own Ads Quality Index, where MiniMax H3 is indexed at 100. The index measures how often a model produces a clip that works as an ad: a clip passes only when prompt following, reference consistency, and visual realism each score 6 out of 10 or higher.

Against its base, Creatify says subject consistency rose from 83.3% to 94.4%, brief completion on difficult production briefs went from 27.8% to 50.0%, and visible defects fell from 1.11 to 0.33 per clip, a 70% reduction.

The company says its automated judges were validated against human preferences and that human studies used anonymized, randomized outputs. Those details are what you would want to see, but the evaluation is still Creatify's own, and no outside party has reproduced the index.

The price is the more disruptive claim

Boreal-H3 is listed at $0.04 per second at 768p. Creatify compares that to Seedance 2.5 at roughly $0.47 per second, which it describes as about twelve times more expensive.

Generation speed is cited at 1.1 seconds per video-second at 768p, meaning a 10-second clip takes about 11 seconds and costs $0.40. Seedance 2.5 is listed at 46.5 seconds per video-second, which would make Boreal-H3 roughly 42 times faster. MiniMax H3 itself ran at 32.3 seconds per video-second in the same comparison.

On a published 15-second product demo, Creatify says Boreal-H3 rendered in 84 seconds against 411 seconds for Seedance 2.5.

There is a useful caveat buried in the comparison. Speed and cost figures depend heavily on resolution and settings, and a 768p output is not the same product as a 1080p one. The right read is that Boreal-H3 is positioned as a high-volume, cost-sensitive tool rather than a quality-first one, and the price reflects that.

Why post-training on ad data is a sensible strategy

The strategic logic here is worth spelling out, because it points at where a lot of video generation is heading.

General-purpose video models are trained on whatever is available on the internet. That produces broad competence and mediocre reliability on any specific task. An advertiser does not need a model that can render a documentary or an anime sequence. It needs a model that reliably produces a clip where the product looks right and the person does not morph mid-sentence.

Post-training on real ad project data narrows the model toward that target. The base capability comes from MiniMax H3, which is already strong, and the specialization comes from data the base lab does not have. Creatify's advantage is that it has spent years running ad projects and knows what fails.

This is a model for how smaller companies can compete without training a foundation model from scratch. It requires a strong open base, a proprietary dataset, and a clear definition of the task. The output is a better generator for one job, rather than a better generator in general.

The math also works in the base lab's favor, which is why MiniMax partnered on it. Every specialized model built on H3 extends the base model's influence without requiring MiniMax to serve every vertical itself. It is the same dynamic that made open-weight models strategically valuable to their publishers: the ecosystem does work the original lab would find unprofitable.

There is a limit worth naming. Specialization narrows the model, and a narrower model is less useful outside its lane. Boreal-H3 will not be the right tool for an indie film, a music video, or a product visualization that needs unusual camera work. That is a feature rather than a bug, provided buyers understand what they are purchasing.

Two control mechanisms

Boreal-H3 gives creative teams two ways to direct a shot.

Keyframe control in image-to-video preserves the subject and visual style of an opening frame, with an optional ending frame to set where the shot lands. That is the mechanism for making a product shot start and end exactly as the brand requires.

Multi-reference control in reference-to-video combines up to nine images, three videos, and three audio clips into a single coherent shot. That is the mechanism for pulling a campaign's existing assets into a new piece of footage without a reshoot.

Text-to-video is supported as well, with camera direction and natural transitions between scenes. For a marketing team, the useful pattern is to lock the brand-critical frames with keyframe control and let the model handle the connective tissue.

What to watch

Three things would confirm whether this matters beyond a press release.

Independent evaluation of the Ads Quality Index. A pass rate that only its creator can measure is a marketing claim until someone else tests it. Creatify's methodology is more careful than most, with human validation and randomized outputs, and the index is still published by the company selling the model.

Real-world resolution. If Boreal-H3 holds up at 1080p rather than just 768p, the twelve-times-cheaper comparison gets much more credible, because the alternative stops being apples-to-apples once you need higher output. Advertisers who need broadcast or large-format placement will not accept 768p regardless of price.

Adoption by advertisers rather than agencies. The tool matters if brands build workflows around it, and that shows up in whether the model's output starts appearing in live campaigns. Agencies testing a tool is a different signal from brands depending on one.

There is a fourth question that applies to the whole category. Post-trained models inherit their base model's limitations, including any licensing restrictions. MiniMax H3 excludes several territories, and Boreal-H3 built on top of it may carry the same constraints. An advertiser operating globally should check territory coverage before committing a campaign pipeline.

Boreal-H3 is aimed at a budget line rather than a benchmark title. On the numbers Creatify published, it has a plausible case, and the case rests less on quality than on the fact that a clip which reliably works as an ad is worth more than a clip that is merely impressive.