Xiao Yang

Research Director @Canva, focusing on Generative AI. Past @ByteDance @Meta

[Andrew Ng @AndrewYNg]

@Mononofu @JensenHuang This is a false equivalence. Everyone has the right to keep their code private. The problem is when someone tries to stop OTHERS from open sourcing.

[Enze Xie @xieenze_jr]

We (@lawrence_cjs, @yuyangzhao_, @shanasaimoe @songhan_mit ) from the SANA team just posted our second blog: "Bet Small to Win Big: Efficient 2K Video Generation via Deeper-compression AutoEncoder, Linear Attention and Two-Stage Refiner"!

Huge thanks to @ltx_model and @Lightricks for their powerful higher-compression VAE and refiner! 🙏

In this post, we dive into how we adapted and integrated LTX's refiner into our SANA framework to build an efficient Two-Stage pipeline. We share the technical details of this (non-trivial) integration and analyze why it delivers stunning 2K video generation with virtually no extra latency or computational overhead compared to our 720p baseline. Bet small to win big! 🚀

Read more: https://t.co/9Yk683ZSnS

Model: https://t.co/OOOJYXTyXR

Demo: https://t.co/S5XQmE9MNx

[Xiao Yang @YangZuoshi]

Also checkout this youtube Short of our amazing lead @Stefano001 introducing this feature at AI Vision, NYC👏👏: https://t.co/m35cgbrKJ6

[Xiao Yang @YangZuoshi]

Happy to see the Canva's magic layer feature launched succesfully! This feature enables you to generate fully editable, layer-wise designs, and I helped making the underlying foundation model 100% faster😁, check out the official product annoucement: https://t.co/EqDuT4elFm

[Xiao Yang @YangZuoshi]

@yucheng_han @wangbudui 嗯，我感觉至少是qwen-image的程度，但在目前的范式上对模型参数量有硬性要求。现在的端上模型工作比如snapgen++，最大好像就到1.6B的DIT参数量；通俗的讲这个参数量在sd1.5到SDXL之间，从模型容量上讲还是太小了

[Xiao Yang @YangZuoshi]

@teortaxesTex Hi, Helios author here, I believe YuanLab is actually a different bunch of people from us haha 😂，Helios is from Peking University and Bytedance, we are not aware of Yuanlab.

[AK @_akhaliq]

Helios

Real Real-Time Long Video Generation Model

paper: https://t.co/ae0ZH4zPzn

[Xiao Yang @YangZuoshi]

Our official report is released on arxiv!
Paper: https://t.co/ES0rfywyRe
code, weights and demo: https://t.co/2Ih3qlMOTA
page: https://t.co/hb4b9Uk4AO

[Xiao Yang @YangZuoshi]

Today we are thrilled to annouse our latest work: Helios, the world's first open-source 14B video model that performs minute-scale, high-quality video synthesis at 19.5 FPS(!) on a single H100 GPU (https://t.co/hb4b9Uk4AO). Trailer: https://t.co/Fx90OFfgSV (1/n)

[Wildminder @wildmindai]

Seedance 2 just dropped as open-source.!.. ok, mostly...

Helios - 14B model for minute-scale vid gen.
- high-quality, fast
-No KV-cache, no quantization, no TinyVAE, just pure architectural efficiency.
- T2V/I2V/V2V ready.

https://t.co/cIUYCOnBGw

[Bin Lin @LinBin46984]

🤯Real-time video generation just got HUGE.
Introducing Helios: A 14B parameter model running at 19.5 FPS on a single H100.
Cheaper, Faster, and Stronger than 1.3B models, generating minute-long videos without the usual tricks (even no quantization).
https://t.co/wdo1L2BbnV

[LMSYS Org @lmsysorg]

🎉 Meet Helios, a 14B video generation model that runs at 19.5 FPS on a single NVIDIA H100 GPU. Day-0 support is now live in SGLang!
🎬 Unified T2V / I2V / V2V: all three tasks in one architecture
⚡️ 19.5 FPS on a single H100 — no KV-Cache, no sparse/linear attention tricks needed
📽️ Minute-long video generation: stable, coherent, no Self-Forcing or Error-Banks required
🧠 14B model trainable at image-model batch sizes — no parallelism or sharding infrastructure needed
Model Page: https://t.co/aYF46REJw2
Related PR: https://t.co/xiaIPXdOkG
Generate now with SGLang!

[Xiao Yang @YangZuoshi]

Finally, here is bonus video showing Helios performing real-time generation, just as an demo of the speed of the model. Hope you enjoy it! 😄https://t.co/iQEdAMPqpv (n/n)

[Xiao Yang @YangZuoshi]

Today we are thrilled to annouse our latest work: Helios, the world's first open-source 14B video model that performs minute-scale, high-quality video synthesis at 19.5 FPS(!) on a single H100 GPU (https://t.co/hb4b9Uk4AO). Trailer: https://t.co/Fx90OFfgSV (1/n)

[Xiao Yang @YangZuoshi]

We also release a detailed report of Helios, including all model information, training techniques, and thorough validations and ablation studies (https://t.co/qSHFOCU3dX) (8/n)