Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Pusa V1.0 Surpassing Wan-i2v-14b With $500 Training Cost By Vectorized Timestep Adaptation

AK

501,816 subscribers

42,170 views • 1 year ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

11 Comments

AK1 year ago

model:

Mobile Scanner1 year ago

Scan any documents, convert images into text, PDF files, etc. 👍

POM1 year ago

“By finetuning the SOTA Wan2.1-T2V-14B model with VTA…”

CRYPT_ EXE1 year ago

Let's summon the goat @Kijaidesign for an fp8 quantization if it's worth it👐

Wyz Borrero - sol/acc1 year ago

All the I2V examples are taken directly from @GoogleLabs FlowTV Veo channels. And they call this "novel", "innovation"...

Maga-Dz1 year ago

Open source as well thank you 👍

JustInEchoes1 year ago

@fofrAI 👀 New open source video gen model

Himanshu Kumar1 year ago

Cheap training doesn't always mean better performance. True value lies in real-world application.

Sebastian1 year ago

Wow thats wild

Dhruv1 year ago

Wow

Safety Marc1 year ago

Looks quite good!

Related Videos

holy moly 🔥 Pusa 1.0 is a new image-to-video finetune based on Wan2.1-T2V-14B that surpasses Wan-I2V-14B and is: > 200x cheaper to train (only cost $500 🤯) > 5x faster inference > supports start-end frames, video extension and video completion

holy moly 🔥 Pusa 1.0 is a new image-to-video finetune based on Wan2.1-T2V-14B that surpasses Wan-I2V-14B and is: > 200x cheaper to train (only cost $500 🤯) > 5x faster inference > supports start-end frames, video extension and video completion

Dreaming Tulpa 🥓👑

15,688 views • 1 year ago

🚀 Just released a LoRA for Wan 2.1 that adds realistic drone-style push-in motion. Model: Wan 2.1 I2V - 14B 720p Trained on 100 clips — and refined over 40+ versions. Trigger: Push-in_camera 🎥 + ComfyUI workflow included for easy use Perfect if you want your videos to actually *move*. 👉 #AI #LoRA #wan21 #generativevideo ComfyUI Made in collaboration with Kartel.ai

🚀 Just released a LoRA for Wan 2.1 that adds realistic drone-style push-in motion. Model: Wan 2.1 I2V - 14B 720p Trained on 100 clips — and refined over 40+ versions. Trigger: Push-in_camera 🎥 + ComfyUI workflow included for easy use Perfect if you want your videos to actually move. 👉 #AI #LoRA #wan21 #generativevideo ComfyUI Made in collaboration with Kartel.ai

Lovis Odin

18,366 views • 1 year ago

Every photo on your phone is now a portal. Lyra 2.0 by NVIDIA - explorable 3D world gen from single images! -long-horizon camera trajectories with 3D consistency; - DMD-distilled 4-step inference; - outputs 3D Gaussians + meshes via Wan 2.1-14B; - minimal drift. Game devs are cooking entire levels from one piece of concept art

Every photo on your phone is now a portal. Lyra 2.0 by NVIDIA - explorable 3D world gen from single images! -long-horizon camera trajectories with 3D consistency; - DMD-distilled 4-step inference; - outputs 3D Gaussians + meshes via Wan 2.1-14B; - minimal drift. Game devs are cooking entire levels from one piece of concept art

Wildminder

20,854 views • 3 months ago

Delhi CM Rekha Gupta distributing water ATM Cards to people on stage with mics and cameras Cost of Card : ₹50 Cost of printing dummy big size card for PR : ₹500 This is how tax payer money is wasted in this country by Anpadhs.

Delhi CM Rekha Gupta distributing water ATM Cards to people on stage with mics and cameras Cost of Card : ₹50 Cost of printing dummy big size card for PR : ₹500 This is how tax payer money is wasted in this country by Anpadhs.

Roshan Rai

280,362 views • 3 months ago

In 1971, Carolyn Davidson, a student at Portland State University, designed the iconic Nike “swoosh” logo for a mere $35. Despite her modest hourly wage of $2, Davidson’s creation would soon become a global symbol of athletic excellence. By 1983, Nike recognized Davidson’s contribution with a gold Swoosh ring and 500 shares of Nike stock. Over the years, the value of these shares has skyrocketed, far surpassing the initial payment. Today, that stock is worth millions, far surpassing the modest initial payment.

VisionaryVoid

63,005 views • 5 months ago

Generalist AI unveiled GEN-0, an embodied foundation model family that hits a 7B-parameter “intelligence threshold”; small models ossify, larger ones keep improving, and is trained on 270,000+ hours of real-world manipulation data growing by ~10,000 hours/week. It introduces Harmonic Reasoning (think+act in continuous time), shows predictable scaling laws, runs across 6–16+ DoF robots, and an infra stack that ingests the equivalent of 6.85 years of experience per training day, promising faster adaptation with less post-training for factories, logistics, and service workflows.

Generalist AI unveiled GEN-0, an embodied foundation model family that hits a 7B-parameter “intelligence threshold”; small models ossify, larger ones keep improving, and is trained on 270,000+ hours of real-world manipulation data growing by ~10,000 hours/week. It introduces Harmonic Reasoning (think+act in continuous time), shows predictable scaling laws, runs across 6–16+ DoF robots, and an infra stack that ingests the equivalent of 6.85 years of experience per training day, promising faster adaptation with less post-training for factories, logistics, and service workflows.

Chubby♨️

21,151 views • 9 months ago

AWS has activated Project Rainier, one of the world’s largest AI compute clusters with nearly 500,000 Trainium2 chips brought online in under a year. Anthropic is already training and serving Claude on the system and is slated to scale past 1 million Trainium2 chips by the end of 2025. The UltraServer/UltraCluster design (NeuronLinks + EFA) speeds massive distributed training, while AWS’s vertically integrated stack and data-center efficiency targets aim to cut cost and environmental impact.

AWS has activated Project Rainier, one of the world’s largest AI compute clusters with nearly 500,000 Trainium2 chips brought online in under a year. Anthropic is already training and serving Claude on the system and is slated to scale past 1 million Trainium2 chips by the end of 2025. The UltraServer/UltraCluster design (NeuronLinks + EFA) speeds massive distributed training, while AWS’s vertically integrated stack and data-center efficiency targets aim to cut cost and environmental impact.

Wes Roth

50,699 views • 9 months ago

Groq founder Jonathan Ross explains why their architecture is better for inference than GPUs. Inference doesn't require as much compute power as training. What matters in inference is speed, cost, and energy consumption. $NVDA was going to lose market share as AI workloads shift to inference. They saw this and made a highly strategic move by striking a deal with Groq. $NVDA is already dominating training, and now it's also well-positioned to dominate inference.

Groq founder Jonathan Ross explains why their architecture is better for inference than GPUs. Inference doesn't require as much compute power as training. What matters in inference is speed, cost, and energy consumption. $NVDA was going to lose market share as AI workloads shift to inference. They saw this and made a highly strategic move by striking a deal with Groq. $NVDA is already dominating training, and now it's also well-positioned to dominate inference.

Oguz Erkan

97,186 views • 7 months ago

🚨Some sources even claim that Persija Jakarta’s budget reaches around Rp540 billion just to support Shin Tae Yong in the transfer market. That figure is quite astonishing, as it is nearly twice the amount spent by Stoke City some time ago to build their Clayton Wood training facility, which “only” cost around Rp230 billion 🤯 With significant fresh funds available, Persija Jakarta has a great opportunity to upgrade their training facilities, but the management seems to have other plans.

🚨Some sources even claim that Persija Jakarta’s budget reaches around Rp540 billion just to support Shin Tae Yong in the transfer market. That figure is quite astonishing, as it is nearly twice the amount spent by Stoke City some time ago to build their Clayton Wood training facility, which “only” cost around Rp230 billion 🤯 With significant fresh funds available, Persija Jakarta has a great opportunity to upgrade their training facilities, but the management seems to have other plans.

The Wall Pass Journal

219,012 views • 2 months ago

Testing image-to-CAD for SR Platform on a new set of GPUs 🔥 We are working with lium.io (SN51) to showcase the efficiency of decentralized compute for SR Platform — from CAD generation to RL training and sim-to-real validation What it means: > Flexible, on-demand rental for us. Pay with crypto. > Diverse GPU options, powered by Openτensor Foundaτion's incentive alignment. > Lower equipment cost → higher margin for the network. Onchain robots, onchain compute.

Testing image-to-CAD for SR Platform on a new set of GPUs 🔥 We are working with lium.io (SN51) to showcase the efficiency of decentralized compute for SR Platform — from CAD generation to RL training and sim-to-real validation What it means: > Flexible, on-demand rental for us. Pay with crypto. > Diverse GPU options, powered by Openτensor Foundaτion's incentive alignment. > Lower equipment cost → higher margin for the network. Onchain robots, onchain compute.

Strike Robot

18,134 views • 2 months ago

Here is another snippet of my conversation with Ryan James Girdusky. There is a societal cost to enforcing the law and that cost is paid by society through danger to both our law enforcement officers and our civilians. This cost is an inevitability built it to the reality of enforcement and human nature. But America’s officers are very well trained and considering the size of our country and the number of officers or civilians lost during enforcement — we’re very successful at keeping this cost to a minimum. American law enforcement leaders are dedicated to a constant and continuing effort to minimize risks to both officers and civilians. Every life lost in an enforcement operation becomes a learning experience for officers throughout the country. Our law enforcement leaders have done a tremendous job developing training and strategies that have rendered use-of-force tragedies into anomalies. Americans should be very proud of where we are today as far as training of our officers.

Here is another snippet of my conversation with Ryan James Girdusky. There is a societal cost to enforcing the law and that cost is paid by society through danger to both our law enforcement officers and our civilians. This cost is an inevitability built it to the reality of enforcement and human nature. But America’s officers are very well trained and considering the size of our country and the number of officers or civilians lost during enforcement — we’re very successful at keeping this cost to a minimum. American law enforcement leaders are dedicated to a constant and continuing effort to minimize risks to both officers and civilians. Every life lost in an enforcement operation becomes a learning experience for officers throughout the country. Our law enforcement leaders have done a tremendous job developing training and strategies that have rendered use-of-force tragedies into anomalies. Americans should be very proud of where we are today as far as training of our officers.

Marina Medvin 🇺🇸

28,927 views • 7 months ago

🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠 Hybrid Linear–Softmax Attention 3 gated linear layers + 1 gated-softmax anchor (75% linear / 25% softmax) 🧱 Block Attention Residuals (AttnRes) Boosts deep-layer effective rank by ~12% 🏗️ Unified 5B & 14B models Trained from scratch under very limited resources: • 5B → only 16 nodes of H100 • 14B → 48 nodes of B200 Results 📊 • 84.30 VBench Total • 3.2× faster DiT forward than matched full-softmax at 720p/60s • 720p/5s in 13.06s on a single H100 with Sol-Engine • 120× faster than Wan 2.2-A14B under the same one-H100 setup SANA-Video 2.0 shows that high-quality 720p video generation can be both efficient and practical on a single GPU. 🎬 Project: 📄 Paper: 💻 Code: Proud of the team! 🎉 More details below 🧵

🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠 Hybrid Linear–Softmax Attention 3 gated linear layers + 1 gated-softmax anchor (75% linear / 25% softmax) 🧱 Block Attention Residuals (AttnRes) Boosts deep-layer effective rank by ~12% 🏗️ Unified 5B & 14B models Trained from scratch under very limited resources: • 5B → only 16 nodes of H100 • 14B → 48 nodes of B200 Results 📊 • 84.30 VBench Total • 3.2× faster DiT forward than matched full-softmax at 720p/60s • 720p/5s in 13.06s on a single H100 with Sol-Engine • 120× faster than Wan 2.2-A14B under the same one-H100 setup SANA-Video 2.0 shows that high-quality 720p video generation can be both efficient and practical on a single GPU. 🎬 Project: 📄 Paper: 💻 Code: Proud of the team! 🎉 More details below 🧵

Enze Xie

25,150 views • 16 days ago

Elon Musk: We hope to achieve full rocket reusability this year, a profound invention that will drop the cost of access to space by a factor of 100. “The key breakthrough that SpaceX is hoping to achieve this year is full reusability. No one has ever achieved full reusability of a rocket, which is very important for the cost of access to space. We've achieved partial reusability with Falcon 9 by landing the boost stage. We've now landed the boost stage over 500 times, but we have to throw away the upper stage. The upper stage burns up on reentry for Falcon 9, and the cost of that is equivalent to a small-to-medium-sized jet. Hopefully this year, we should prove full reusability for Starship, which will be a profound invention because the cost of access to space will drop by a factor of 100 when you achieve full reusability.” Interview at WEF, January 22, 2026

Elon Musk: We hope to achieve full rocket reusability this year, a profound invention that will drop the cost of access to space by a factor of 100. “The key breakthrough that SpaceX is hoping to achieve this year is full reusability. No one has ever achieved full reusability of a rocket, which is very important for the cost of access to space. We've achieved partial reusability with Falcon 9 by landing the boost stage. We've now landed the boost stage over 500 times, but we have to throw away the upper stage. The upper stage burns up on reentry for Falcon 9, and the cost of that is equivalent to a small-to-medium-sized jet. Hopefully this year, we should prove full reusability for Starship, which will be a profound invention because the cost of access to space will drop by a factor of 100 when you achieve full reusability.” Interview at WEF, January 22, 2026

ELON CLIPS

34,121 views • 2 months ago

VIDEO: We Are Hungry, Protesters Decry Economic Hardship Protesters have stormed the streets of Lagos to express their grievances over the current economic hardship. They are also protesting the government’s alleged failure to implement agreements reached between both parties on October 2, 2023, following the removal of fuel subsidy. One of the protesters in a video shared by our correspondent was heard saying, “Hungry wan kill us for this country, make una comot Tinubu, we no want Tinubu again. “Him government no favor us, him own time too dey difficult. Hungry dey kill us with our children. Fuel don cost, gas don cost, all of them don cost. ` Another protester was heard saying “Pure water na N50 , we are hungry o, attachment N5,000 ” Credit: Olabode Olalekan --- Follow us for more breaking news and videos

VIDEO: We Are Hungry, Protesters Decry Economic Hardship Protesters have stormed the streets of Lagos to express their grievances over the current economic hardship. They are also protesting the government’s alleged failure to implement agreements reached between both parties on October 2, 2023, following the removal of fuel subsidy. One of the protesters in a video shared by our correspondent was heard saying, “Hungry wan kill us for this country, make una comot Tinubu, we no want Tinubu again. “Him government no favor us, him own time too dey difficult. Hungry dey kill us with our children. Fuel don cost, gas don cost, all of them don cost. ` Another protester was heard saying “Pure water na N50 , we are hungry o, attachment N5,000 ” Credit: Olabode Olalekan --- Follow us for more breaking news and videos

Punch Newspapers

134,731 views • 2 years ago

🧠💡 What if your 7B model could beat GPT-4o and Qwen2.5-72B—using just 11k training samples? No distillation. No warm-start. Just smart data and reinforcement learning. Inspired by Moravec’s Paradox, we let the model decide what's actually hard. 🚨 New paper: "SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement" We show how ThinkLite-VL-7B achieves SoTA on MathVista—75.1%, surpassing much larger models. 👇 Here’s how we did it: 🔗 🧠 Code: #AI #VisionLanguageModels #ReinforcementLearning #MachineLearning #LessIsMore

🧠💡 What if your 7B model could beat GPT-4o and Qwen2.5-72B—using just 11k training samples? No distillation. No warm-start. Just smart data and reinforcement learning. Inspired by Moravec’s Paradox, we let the model decide what's actually hard. 🚨 New paper: "SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement" We show how ThinkLite-VL-7B achieves SoTA on MathVista—75.1%, surpassing much larger models. 👇 Here’s how we did it: 🔗 🧠 Code: #AI #VisionLanguageModels #ReinforcementLearning #MachineLearning #LessIsMore

Furong Huang

63,321 views • 1 year ago

Check out Robuild's autonomous construction robots They recently completed a 500 ft² LVP flooring installation, with 600 install planned by 2026. A few tech details: • Labour shortages & cost savings are driving developers to look at robotics • The system works in busy, varied houses and uses onboard vision to align and place planks without pre-mapping. • It's built on ROS, using NVIDIA Omniverse, Isaac Gym, and custom PID for precision.

Check out Robuild's autonomous construction robots They recently completed a 500 ft² LVP flooring installation, with 600 install planned by 2026. A few tech details: • Labour shortages & cost savings are driving developers to look at robotics • The system works in busy, varied houses and uses onboard vision to align and place planks without pre-mapping. • It's built on ROS, using NVIDIA Omniverse, Isaac Gym, and custom PID for precision.

Jack 🤖

223,954 views • 1 year ago

Germany's hydrogen train flagship became a taxpayer-funded shambles. The Taunus Network ordered 27 Alstom hydrogen trains at a cost of roughly €500 million. The pitch was simple: Replace diesel, cut emissions, build the world's largest hydrogen train fleet and save the planet. Then reality hit: The trains failed. By early 2025, 18 of 24 hydrogen trains had to be replaced by diesel units. All taxpayers got for their €500 million were delays, replacement buses and service cuts. The "clean rail revolution" ended with diesel coming to the rescue. As if it needed it, Germany just got another expensive lesson in subsidized green fantasy.

Germany's hydrogen train flagship became a taxpayer-funded shambles. The Taunus Network ordered 27 Alstom hydrogen trains at a cost of roughly €500 million. The pitch was simple: Replace diesel, cut emissions, build the world's largest hydrogen train fleet and save the planet. Then reality hit: The trains failed. By early 2025, 18 of 24 hydrogen trains had to be replaced by diesel units. All taxpayers got for their €500 million were delays, replacement buses and service cuts. The "clean rail revolution" ended with diesel coming to the rescue. As if it needed it, Germany just got another expensive lesson in subsidized green fantasy.

Electroverse

19,008 views • 2 months ago

𝐃𝐌𝐅 𝐒𝐮𝐧𝐝𝐚𝐫𝐠𝐚𝐫𝐡 𝐟𝐮𝐞𝐥𝐬 𝐭𝐡𝐞 𝐝𝐫𝐞𝐚𝐦𝐬 𝐨𝐟 𝐫𝐢𝐬𝐢𝐧𝐠 𝐇𝐨𝐜𝐤𝐞𝐲 𝐬𝐭𝐚𝐫𝐬 Sundargarh, Odisha, home to hockey legends like Michael Kindo and Dilip Tirkey, continues to nurture India’s next generation of champions. With dedicated support from the District Mineral Foundation (DMF) by the Ministry of Mines, young athletes are receiving world-class training and opportunities to excel. Key interventions by DMF include: ✅ Construction of 16 state-of-the-art hockey training centres with modern facilities, built at a cost of ₹12 crore ✅ Free hockey coaching for 1,000 students, helping them access professional guidance ✅ Continued efforts to strengthen sports infrastructure and talent development across the district These initiatives are empowering Sundargarh’s budding players to transform their passion into national and international success. Ministry of Mines dmfgov

𝐃𝐌𝐅 𝐒𝐮𝐧𝐝𝐚𝐫𝐠𝐚𝐫𝐡 𝐟𝐮𝐞𝐥𝐬 𝐭𝐡𝐞 𝐝𝐫𝐞𝐚𝐦𝐬 𝐨𝐟 𝐫𝐢𝐬𝐢𝐧𝐠 𝐇𝐨𝐜𝐤𝐞𝐲 𝐬𝐭𝐚𝐫𝐬 Sundargarh, Odisha, home to hockey legends like Michael Kindo and Dilip Tirkey, continues to nurture India’s next generation of champions. With dedicated support from the District Mineral Foundation (DMF) by the Ministry of Mines, young athletes are receiving world-class training and opportunities to excel. Key interventions by DMF include: ✅ Construction of 16 state-of-the-art hockey training centres with modern facilities, built at a cost of ₹12 crore ✅ Free hockey coaching for 1,000 students, helping them access professional guidance ✅ Continued efforts to strengthen sports infrastructure and talent development across the district These initiatives are empowering Sundargarh’s budding players to transform their passion into national and international success. Ministry of Mines dmfgov

G Kishan Reddy

11,230 views • 8 months ago

Enough with LLMs - exciting things are happening in the world of atoms. This is Stanford ALOHA, a low-cost and agile robot platform. The whole system is open-source (!!): hardware design, CAD models for 3D printing, simulator, and training code. Time to graft a physical arm onto GPTs 🦾 Led by my friend Tony Zhao at Stanford, and advised by Chelsea Finn Sergey Levine Vikash Kumar. Project page: If you don't want to 3D print your own components, you can buy the setup at

Enough with LLMs - exciting things are happening in the world of atoms. This is Stanford ALOHA, a low-cost and agile robot platform. The whole system is open-source (!!): hardware design, CAD models for 3D printing, simulator, and training code. Time to graft a physical arm onto GPTs 🦾 Led by my friend Tony Zhao at Stanford, and advised by Chelsea Finn Sergey Levine Vikash Kumar. Project page: If you don't want to 3D print your own components, you can buy the setup at

Jim Fan

818,539 views • 3 years ago