Пропустить навигацию
Introducing Boreal-H3 — a video model built for ads and our next step toward recursive…
0:00 / 0:45

Introducing Boreal-H3 — a video model built for ads and our next step toward recursive…

Creatify Labs

@Creatify_Labs2 тыс. подписчиков

Открыть в X
205 отметок «Нравится»СкачатьОригинальный пост

775 273 просмотра

Introducing Boreal-H3 — a video model built for ads and our next step toward recursive self-improvement in video generation. A good-looking video isn’t enough. The product has to stay the same. The actor has to stay the same. The label has to be right. And the action in the brief actually has to happen. So we post-trained MiniMax H3 specifically for advertising. But this isn’t a one-off SFT or LoRA fine-tune. We built a closed-loop system that learns what to improve next. Human-calibrated evaluation diagnoses failures and guides the next intervention: targeted data collection, reinforcement learning, or inference optimization. When the feedback is unreliable, we revise the evaluator or reward—not just the generator. Every experiment feeds into shared memory, informing the next training decision. The model improves, and so does the process that produces its successor. The results: → 85.3% reference fidelity — highest among the frontier video generation models we evaluated → Brief success: 28% → 50% → Identity match: 83% → 94% → Visible defects per clip: down 70% → Generation time and estimated cost: down 20% Boreal-H3 doesn’t just make better-looking video. It makes more usable ads. Credit to the MiniMax (official) team for the foundation we’re building on. This launch is a checkpoint, not the finish line. We’re building more than a better video model. We’re building a system that learns how to make the next one better.

71 комментарий

Комментарии из оригинального поста в X (Twitter)

  • Фото профиля Creatify Labs

    @Creatify_Labs

    1/ One more thing about the launch video above: It was created by our researchers using our new Ad Agent, powered by Boreal-H3. Boreal-H3 gives the agent a video model optimized for the things ads actually require: product fidelity, creator consistency, directability and brief execution. Our researchers used the ad ggent to make the launch video. Now we want to see what you make with it. → Try the new Creatify Ad Agent: → Try Boreal-H3: → Technical report:

    16 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    2/ The capability we care about most for advertising: reference fidelity. A product can’t become a slightly different product halfway through the shot. The shape, packaging, label and visual identity have to survive motion. Boreal-H3 achieves 85.3% reference fidelity — the highest among the frontier video generation models we evaluated. Look closely at the product↓

    14 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    3/ But an ad is more than one faithful frame. The model also has to follow the actual creative brief across shots, actions and transitions. For example: cliff jump → wingsuit flight → beach landing → vendor → final bite. Boreal-H3 completes the full sequence. Following the brief means reaching the payoff.

    15 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    4/ The visual examples show the difference. The evaluation tells the same story. → 85.3% reference fidelity — highest among the frontier models evaluated → 133.3 Ads Quality index — highest observed across six frontier models Ads Quality jointly measures reference consistency, prompt following and visual realism. We optimize for whether a generation is actually usable as an ad—not just whether it looks good.

    17 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    5/ So what did we actually change? This is not a one-off SFT or LoRA fine-tune. Our post-training system closes the loop: real-use feedback → capability diagnosis → choose what to improve → data / training / inference intervention → evaluate against the incumbent → record the result → repeat SFT and RL are tools inside the loop—not the whole recipe.

    13 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    6/ A concrete example: If the model starts drifting on product packaging, the answer isn’t simply “train more.” The system can identify the failure, collect more packaging and product-handling data, curate it, train a new candidate, and test whether fidelity improves without regressing elsewhere. And if the evaluator itself is noisy, we fix the evaluator first. Every experiment becomes memory for the next intervention. That’s what we mean by toward recursive self-improvement in video generation.

    12 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    Try Boreal-H3: Try Ad Agent:

    14 отметок «Нравится»

  • Фото профиля Jaymin Shah

    @JayminSOfficial

    Reference fidelity and brief success are exactly the metrics that turn video models from impressive demos into usable ad infrastructure!

    15 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    Exactly. A demo can look great and still be useless for an ad. Those are the numbers that matter.

    2 отметки «Нравится»

  • Фото профиля Tiffany Fong

    @TiffanyFong

    Insane 🔥

    7 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    Thanks Tiffany! 🙌

  • Фото профиля Not Elon Musk

    @ElonMuskAOC

    It’s getting too easy now!

    6 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    That's kind of the point 😄

    3 отметки «Нравится»

  • Фото профиля Paul Mit

    @pmitu

    wow what a movie check my result:

    6 отметок «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    In one shot is wild, Paul. The UI stayed clean all the way through. Love this!

    2 отметки «Нравится»

  • Фото профиля 🚨 AI News | TestingCatalog

    @testingcatalog

    Quality / price ratio is at peak 💪👀

    3 отметки «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    Indeed! We offer the best quality / price for ads for sure!

    1 отметка «Нравится»

  • Фото профиля 𝚠𝚘𝚕𝚏⚡️

    @elong_tusks

    no one absolutely no one gives a fuck 🖕🏼🤣

    3 отметки «Нравится»

  • Фото профиля Jeddi

    @antinertia

    love it! how long did it take to build this new version?

    3 отметки «Нравится»

  • Фото профиля Creatify Labs

    @Creatify_Labs

    Thanks! The feedback loop behind it was months of work, and it keeps improving the model from here.

    2 отметки «Нравится»