One image + text + camera trajectory = controllable worlds. All on a single GPU. Our research team just released SANA-WM, a 2.6B open source world model natively trained for 60-second video generation with precise camera control.

The architecture features Hybrid Linear Attention, Dual-Branch Camera Control, Two-Stage Generation Pipeline, and a Robust Annotation Pipeline. All of this combines for stronger action-following accuracy and higher throughput while maintaining visual quality.

You can find the full project including the paper, code, and model here 👇

this sounds super cool! creating worlds with just a single gpu is a game changer. can't wait to see what people do with SANA-WM!

60 seconds of video. One image to start. One GPU to run it. Open source. The game engine just became a prompt.

Looks great. Can you link the paper and repo, name training data sources/licenses, and state the exact GPU/VRAM for the single GPU claim? That would let others reproduce the camera control results.

2.6B for 60-second world-consistent video with camera control is a wild benchmark. The open-source release is what makes this actually useful — most teams don't need Hollywood quality, they need something they can run and fine-tune. What hardware are people actually running inference on?

love how you're pushing the boundaries of video generation with SANA-WM

@llmgram

You're on the right track Nvidia, but please make it more performant and keep training. Go full send on compute for training.

SANA-WM 2.6B open source world model for 60-second video on a single GPU

Looking very soulless, generic and not worth it!

Can it run on a 4090?

$NVDA seriously is going to rule the world

One image + text + trajectory = controllable worlds on a single GPU. Simulation is becoming cheap and precise.

bbox queen looks clean, seedance 2 with gpt image 2, huh? got a side by side without seedance to compare?

Love this

"single GPU" is the part that changes everything here. world models for video have always felt like a data center problem, shrinking that down to one GPU opens this up to a completely different audience

SANA-WM is concrete enough to test. One image, text, camera path, and single GPU are the setup. Control drift is the receipt.

2.6B for 60-second generation is compelling if SANA-WM keeps temporal drift under control.

@KashJim50723 🤷‍♂️

Which GPU, and how much VRAM?

Hey @NVIDIAAIDev — got SANA-WM running on one H100. "Single GPU" holds at runtime, but the install path is rough: HF repo has no model_index.json, real setup needs vendoring the feat/sana-wm PR branch + mmcv 1.7.2 + flash-attn stub. Any chance of an official quickstart? 🙏

What's the max resolution you folks are targeting for high-res video?

Awesome share

Oh I wanna test this

One image plus text plus camera trajectory and you get a controllable world. That formula is deceptively simple for what it produces

wao great now give me 100 B200 for my startup.

Wuuut 👀😳🤯

Interesting how "video generation" is quietly evolving into controllable spatial simulation.

thats awesome! pretty cool!

what's the latency on a consumer GPU though?

Finally a world model that doesn’t require a datacenter to test your camera moves

means i have to huy it ASAP🥹