Co-Founder & Chief Scientist @ Creatify AI ICML Test-of-Time Award Co-author of DPR (RAG-core) Past: FAIR @MetaAI
Joined May 2020
- Tweets90
- Following308
- Followers1.1K
- Likes177
Pinned Tweet
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/bo…
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.
A key lesson from our work on Boreal-H3: evaluation needs to reflect the domain we're optimizing for.
Generic video quality benchmarks don't fully capture what matters in advertising—product fidelity, identity consistency, and brief following.
Domain-specific evaluation is essential for both measuring progress and guiding post-training.
Built for ads, scored on ads.
Boreal-H3 tops our Ads Quality Index at 133.3 (MiniMax H3 = 100).
Creatify internal evaluation, Sep 2026, same briefs for every model.
Full table and method → labs.creatify.ai/research/bo…
111
Better quality at lower cost. That's what we're working toward with Boreal-H3.
Beyond post-training, our team has made significant progress on inference optimization to make high-quality video generation more accessible and scalable.
More technical details coming soon.
Better and cheaper than its own base.
Boreal-H3 is the only model in the top-left: 133.3 on ads quality vs. MiniMax H3's 100, at half the price per second.
Read more: labs.creatify.ai/research/bo…
1
1
115
$1 buys twenty five-second video clips with Boreal.
That opens up an exciting direction for creative agents: generate multiple candidates, judge product fidelity and creator realism, then iterate on the strongest ideas.
Our team at Creatify built Boreal through a recursively self-improving post-training loop focused on advertising:
• 81% preference over its base model in decisive blind human comparisons
• Results comparable with frontier models on selected perceptual-quality measures
• $0.01/sec, with a five-second clip generated in ~5 seconds
As a researcher, I’m excited about what happens when generation becomes fast and affordable enough to sit inside an agent’s exploration loop.
Proud of the team bringing this to life: labs.creatify.ai
Introducing Boreal from Creatify Labs. A new text-and-image-to-video model built for video generation. At just one cent per second, Boreal is up to 47x cheaper than Seedance 2.5.
Boreal is an open video base we post-trained on a corpus of real ad footage, creator-style UGC, and production customer briefs.
In blind human review of over 40 production cases, post-training raises Boreal's win rate over the base model to 81%.
We built Boreal inside Creatify for our own users and tested it at scale on real workloads.
Today we're releasing it to everyone.
$1 buys twenty five-second video clips with Boreal.
That opens up an exciting direction for creative agents: generate multiple candidates, judge product fidelity and creator realism, then iterate on the strongest ideas.
Our team at Creatify built Boreal through a recursively self-improving post-training loop focused on advertising:
• 81% preference over its base model in decisive blind human comparisons
• Results comparable with frontier models on selected perceptual-quality measures
• $0.01/sec, with a five-second clip generated in ~5 seconds
As a researcher, I’m excited about what happens when generation becomes fast and affordable enough to sit inside an agent’s exploration loop.
Proud of the team bringing this to life: labs.creatify.ai
This quoted post is unavailable.
Excited to share Creatify Agent: our step toward AI-native ad orchestration.
Creatify Agent is a multi-agent system for ad video and image production. A central producer orchestrates six stages — brief understanding, web research, parallel scene generation, scored audio, layered assembly, and per-scene QA — with a vision-capable critic feeding verdicts back into the loop.
Alongside the system, we're releasing VideoAdAgent Bench — the first public benchmark for AI ad-video production. Creatify Agent ranks first on the composite, hallucination, and pairwise-preference metrics that govern whether a finished ad is shippable, outperforming general-purpose video agents and frontier T2V models.
The design choices we think matter:
→ Reference-anchored generation. The producer can pull 8–10 high-performing ads from TikTok / Instagram in the brand's category before any scene is rendered. URLs and competitor reels enter the same typed asset graph as uploaded assets.
→ Persistent per-brand memory across sessions. Every T2V model and general-purpose video agent we benchmarked starts from zero on every call.
→ In-the-loop hallucination control. The critic runs inside production, not as a post-hoc filter. Each scene is judged against an explicit rubric (product fidelity, object hallucination, text quality, physical violations, anatomy), and failures route back with a structured verdict — which dimension, by how much, on which frame — for surgical per-scene regen.
Try it for free: app.creatify.ai/creative-age…
Full tech report: creatify.ai/research/agent
VideoAdAgentBench: creatify.ai/research/agent-b…
We built Ad Flow because we kept hearing the same thing from performance marketers: the creative process is fragmented across 5+ tools before a single ad goes live.
One canvas. Idea to finished ad. Swap the hook, ship every version.
This is what creative velocity actually looks like.
#AIAdvertising #CreativeAutomation
Every ad tool we've used has had the same problem, constant context switching. Film here, edit there, add audio somewhere else. Creative momentum dies before you can even finish the ad.
Ad Flow fixes that.
Idea to finished ad, without ever leaving the canvas. When something works, swap the hook and ship every version at lightning speed.
Other tools generate pieces. Ad Flow creates the whole picture.
To kick off this launch, we're giving away 100+ Ad Flow Ad templates with replay across different product categories.
Comment "FLOW" to unlock access to the Ad templates with replay.
Make sure to check your Dm's
👉 app.creatify.ai/tool/adflow-…
3
274
Aurora is now live on @fal - ultra-realistic, expressive avatars from a single image + voice.
Built to scale, built to feel real. 🚀
fal.ai/models/fal-ai/creatif…
🚨 Creatify Aurora is now live on fal!
🎭 State-of-the-art image-to-avatar model - ultra-realistic avatars from a single image and voice
✨ Avatars that move, speak, and feel real with exceptional expression
🎨 Versatile across avatar styles and scenarios with scalable, consistent quality
1
6
881
Excited to share our ad clone launch — another brick in the foundation of end-to-end AI creative automation. More to come. ✨
Marketers have been cloning ads forever.
They just never had a button for it.
So… we built one. 🚀
Introducing Ad Clone:
→ Pick any top performing ad from our AI Ad Library
→ Swap in your product
→ Generate a new version built on the same winning structure
(angles, pacing, hooks, CTAs all rebuilt for your brand)
One early tester cloned a competitor’s ad style
Cut their CPA by 50% in 7 days.
Creative isn’t luck anymore.
It’s leverage.
Ad Clone launches today.
Like + RT
Comment “Clone” We’re giving away up to 50,000 free Credits for (Next 9 hours🔥)
1
2
385
Ledell Wu retweeted
We just shipped Sora 2 in our Asset Generator… and it’s killer when combined with other models. Especially since you can’t use humans as reference images for Sora.
It’s crucial for DTC brands & agencies who need video content (esp UGC) at scale
Here's the workflow we’re using for our biggest clients:
→ Product isolation from Nanobanana
→ Creatify’s brand kits to save crucial selling points, unique differentiators → Threaded asset generation to chain multiple outputs and models at the same time
Now for the Sora part:
→ Use the reference image + brand kit to generate UGC videos
→ Ensure the hook and CTA come across in the prompt (dialogue, SFX, ambiance synced)
→ Generate multiple batches
→ Combine with AI Aurora avatars for longer UGC scenes as well
100% all generated in Creatify and zero post-production required
Want to get access?
→Comment "SORA" →Like this post/Follow me
And I'll send the workflow (must be following so I can DM)
Super Excited to introduce Aurora (Audio-driven Ultra-Realistic Rendering of Reactive Avatars) – our latest technical breakthrough: a new state-of-the-art AI foundation model for avatar video generation.
From a single image + an audio clip, Aurora produces a photorealistic video of a person speaking or singing — with full-body expressiveness, not just lip sync.
Aurora is a diffusion-based multimodal model that fuses audio, image, and optional text inputs to generate temporally coherent, emotionally aligned motion across the face, eyes, hands, and overall posture.
See demos + full detailed blog post here: 🔗 creatify.ai/introducing-auro…
What makes this powerful:
- Zero-shot generation with highest visual fidelity
Identity consistency over long videos
- Natural gestures, facial nuance, and subtle motion
- Robust performance across diverse scenarios
It’s one of the most natural, expressive and controllable avatar models we’ve seen — and I’m incredibly proud of my team for making it happen.
🆕 Today we’re introducing Aurora - the state-of-the-art image to avatar model.
From a single image and a voice, it generates ultra-realistic avatars that move, speak, and feel real.
With Aurora, you can scale your most creative marketing ideas. Aurora avatars don’t just say lines, they express them, making your ideas come alive on screen.
Comment "Aurora" & RT to get early access today.
We are launching Design Your Own Avatar 2.0 today.
With DYOA 2.0, you can freely create avatars and customize them in a variety of ways.
→ Outfits that match your style.
→ Backgrounds that set the perfect scene.
→ Camera angles & branding elements all in your control.
👉Try it for unlimited customization: creatify.ai/features/dyoa
Introducing DYOA 2.0 — your new way to make avatars.
This isn’t just an update. It’s a complete reinvention of how brands build digital identities.
Unlimited customization.
10-minute turnaround.
One avatar, endless looks.
No more bottlenecks. No more “almost there.” Just fast, flexible storytelling at scale.
Start creating: creatify.ai/features/dyoa
2
5
736
We’re just getting started. Huge thanks to @stevejang for leading our Series A and sharing our ambition to reshape advertising. Let’s go! 🙌
Psyched to share at @KindredVentures that we co-led a $15.5M investment into @Creatify_AI, which is building at the intersection of generative media, AI agents, and the $1.1 trillion global advertising market!
Video is the most effective medium in digital advertising, but one of the hardest to scale. Production is slow, iteration cycles are expensive, and performance is hit-or-miss with fuzzy feedback loops.
Creatify is compressing what’s traditionally been a fragmented, manual, and expensive process into a fast, generative ad-creative AI agent with built-in testing and feedback loops. Creatify's AdMax is an AI agent designed to not only make video ads, but manage the entire video ad lifecycle.
The founders, @nyn531, @LedellWu, and Chaz Zhou, bring deep product engineering and AI research experience from @Meta, @Snapchat, and FAIR (fka Facebook AI Research). In a short amount of time since launch, they have reached over $9M in ARR and many happy customers including thousands of creators and brands. Happy to invest and support, alongside @WndrCoLLC and prior investors @SamsungNext, @natfriedman, @danielgross and angels. More details in our blog post below 👇🏽👇🏽👇🏽
kindredventures.com/announce…
3
808
We just announced our $15.5M Series A @creatify_ai , co-lead by @WndrCoLLC and @KindredVentures.
Today, we’re unveiling the next chapter of Creatify: AdMax—an autonomous AI agent for full-funnel advertising. 🚀
It’s not just gen AI for creatives. It’s a decision-making system that closes the loop between creative generation, testing, and performance optimization.
AdMax:
🧠 Ingests competitive ad data
🎨 Uses LLMs + VLMs to generate 100s of tailored video variants
📊 Runs live multi-variate tests (ROAS, CPA)
🔁 Analyzes performance feedback to generate smarter creatives & refine targeting
A full-stack agent that learns, adapts, and scales—autonomously.
The future of performance marketing is here → creatify.ai/admax
Huge thanks to our customers: we couldn't reach this far with all your support, feedback, and love!
More details on our blog: creatify.ai/blog/announcing-…
#AdMax #AIagents #SeriesA #Creatify #growthtech #autonomousads
Big news from the Creatify team:
We’ve raised a $15.5M Series A, co-led by @WndrCoLLCand & @KindredVentures
@nyn531 @LedellWu
In just 18 months, we grew to 1.5M users and hit $9M+ ARR.
But the journey taught us something deeper:
→ Making ads is easy
→ Knowing what works is everything
That’s why we built AdMax — the first AI ad agent.
It listens to the market.
It creates. It tests. It learns.
This is just the start of a new era.
Try AdMax → Creatify.ai
Today we are launching Creatify Product Video! Check it out here 👉 creatify.ai/features/product…
More examples in thread & website.
today we launch Creatify Product Video, our newest video ad generator.
this is your high-end product photographer—just drop in your product image and get a cinematic product video clip in return.
it can also generate a UGC-style creator holding your product, narrating with our hyperrealistic lip-sync model—ready in minutes.
i know there are a lot of product video generators out there, but this one is really good. it can save you hours/days, and thousands of dollars.
8
764
Ledell Wu retweeted
RIP influencers.
I created an AI influencer in 60 seconds, and it looks 100% real.
Here’s how I did it (step-by-step):
Excited and super impressed by Deepseek's V3, R1, Janus model release (3 in a row)!
This is a clear win of the open source community and every builders can benefit from the new innovations.
For AI applications this opens up vast new possibilities, enabling significant innovations.
#deepseek #AGI #AIapp #creatify