Twitxy Saved

Ledell Wu

Verified account Creatify AI

@LedellWu

Co-Founder & Chief Scientist @ Creatify AI
ICML Test-of-Time Award
Co-author of DPR (RAG-core)
Past: FAIR @MetaAI
312 Following 1.1K Followers 95 Posts

Latest posts 5

Pinned post

Ledell Wu Verified account @LedellWu
Really excited to finally share Boreal-H3, our video generation model built for advertising.

What excites me most isn’t just the model quality—it’s our post-training approach behind it.

We started with a strong foundation model, then built a feedback-driven post-training system around the things that actually matter for advertising: reference fidelity, brief following, identity consistency, motion, and failure detection.

Evaluation feeds back into what data we collect, which failures we target, and how we design the next round of post-training—so each cycle improves not only the model, but the process for improving it.

The result: Boreal-H3 reaches 85.3% reference fidelity while making major gains in ad usability.

This is a strong example of how much capability can still be unlocked through post-training—and a step toward a more systematic, self-improving paradigm for video generation.

Huge credit to the team behind this. More technical details in the launch post 👇

Read our technical report: labs.creatify.ai/research/borea…
Creatify Labs Verified account @Creatify_Labs Sep 30
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_AI 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.
View quoted post on X
Ledell Wu Verified account @LedellWu
Excited to introduce Max. We’re bringing creative generation and campaign execution together in one AI teammate—so you can spend more time on strategy while Max handles the day-to-day. You stay in control.
Yinan (Steven) Na Verified account @nyn531 19h
Meet Creatify Max, the newest member of your ad team

Everyone gets their own Max. It watches your accounts 24/7, writes the briefs, makes the videos, and launches the campaigns

Max does the work. You make the call.
View quoted post on X
1 replies 1 reposts 15 likes 606 views View on X
Ledell Wu Verified account @LedellWu
Cold starts are easy to overlook until you’re serving video models in production. Getting to the first denoising step in ~7s means much less GPU time spent waiting.

Proud of our team for building this—and excited to share the runtime and what we learned with the community.
Creatify Labs Verified account @Creatify_Labs 4d
We’re open-sourcing Native Video Inference: the Rust + CUDA runtime we built at @Creatify_AI to serve video models.

Cold start to first denoising step: ~7s on a fresh B200, vs 65–75s on @sgl_project Diffusion and vLLM-Omni. Same output, bit for bit.

No PyTorch at inference. Kernels compiled ahead of time. Everything loads in parallel.

On our production traffic, that means 84% less idle GPU time.

Supports MiniMax H3, Wan 2.1 and LTX 2.5. Code and write-up below 👇
View quoted post on X
Ledell Wu Verified account @LedellWu
We’re open-sourcing Native Video Inference: the Rust + CUDA runtime we built at @Creatify_AI to serve video models.

Cold start to first denoising step: ~7s on a fresh B200, vs 65–75s on @sgl_project Diffusion and vLLM-Omni. Same output, bit for bit.

No PyTorch at inference. Kernels compiled ahead of time. Everything loads in parallel.

On our production traffic, that means 84% less idle GPU time.

Supports MiniMax H3, Wan 2.1 and LTX 2.5. Code and write-up below 👇
6 replies 3 reposts 15 likes 1.2K views View on X