Cosmos3-Super-Image2Video · May 31 2026

Cosmos 3 Super AI Video Generator

Cosmos 3 Super AI Video Generator runs entirely in your browser — no GPU, no install. Powered by NVIDIA's open-weights Cosmos3-Super-Image2Video (64B), it's the #1 ranked model on Physics-IQ for turning still images into physics-accurate video clips. Upload one image, describe the motion, and get a production-ready MP4.

See Example Outputs
No GPU requiredCommercial use allowed1–7 second outputOpenMDW-1.1 license
Model
64B
Cosmos3-Super-Image2Video open weights
Physics-IQ Rank
#1
Top open-weights model on Physics-IQ benchmark
Max Duration
7s
1 to 7 seconds
Aspect Ratios
5
16:9 · 9:16 · 1:1 · 4:3 · 3:4 native output

What Is the Cosmos 3 Super AI Video Generator?

Cosmos 3 Super Image to Video is a post-trained specialist model — officially designated Cosmos3-Super-Image2Video — derived from NVIDIA's Cosmos 3 Super foundation model. Released on May 31, 2026 at GTC Taipei, it generates temporally coherent, physics-real video sequences from a single reference image and a natural-language motion prompt.

The combined 64B architecture ranks #1 on Physics-IQ among open-weights image-to-video models — built for creators who need motion that looks like real footage, not synthetic interpolation.

Most AI video tools just interpolate frames. Cosmos 3 Super actually thinks about the scene first — figuring out how objects weigh, how fabric falls, how light behaves — through its 32B Reasoner tower before the 32B Generator renders a single frame. The result looks like real footage, not the uncanny-valley smoothness you get from most AI video.

The model ships under the NVIDIA OpenMDW-1.1 license, which permits commercial use — paid advertisements, client deliverables, content sold on platforms — with a "Built on NVIDIA Cosmos" attribution requirement.

Release date
May 31, 2026 — GTC Taipei / Computex 2026
Architecture
Mixture-of-Transformers (MoT) — 32B Reasoner + 32B Generator
Benchmarks
#1 open-weights model on Physics-IQ, PAI-Bench, and R-Bench
Training data
1.3B+ data points across text, image, video, and audio

Technical Specifications

Model VariantCosmos3-Super-Image2Video
Post-trained for image-to-video generation
Parameters64B
32B Reasoner + 32B Generator, open weights on HF
Physics-IQ Rank#1
Top open-weights model on Physics-IQ benchmark
Output Duration1–7s
2 credits/sec (480p) · 4 credits/sec (720p) — see pricing
Input FormatsJPG · PNG · WEBP
RGB 8-bit · 256p / 480p / 720p supported
LicenseOpenMDW-1.1
Commercial use included with attribution

Cosmos 3 Super AI Video Generator — Real Output Examples

Six clips generated on this site using Cosmos3-Super-Image2Video. No post-processing, no splice cuts. Copy any prompt to use it directly in the generator above.

What Makes Cosmos 3 Super AI Video Generator Different

Six capabilities that separate Cosmos3-Super-Image2Video from other image-to-video models — backed by the underlying architecture, not marketing copy.

Physics-Real Motion Simulation

The Reasoner tower analyzes the input image for material properties, gravity context, and object relationships before the Generator renders a single frame. Water behaves like water. Fabric drapes, not glitches. Ranked #1 on the open Physics-IQ benchmark among open-weights models.

Image-Guided Video Generation

Cosmos3-Super-Image2Video is purpose-trained for image input. Your subject, color palette, and composition are preserved across every frame. The model animates what's there — it doesn't hallucinate new elements from a text description alone.

Prompt-Based Motion Control

Write what you want to happen — push in on the subject, pan left, add wind — and the model does it. No timeline editor, no keyframes. Negative prompts steer away from artifacts when you need tighter control.

Ambient Sound Generated Automatically

Cosmos 3 Super is an omnimodal model — audio is rendered in the same forward pass as the video. Wind noise matches the visible breeze. Water sounds match the splash on-screen. No separate audio tool, no timeline sync required.

Five Platform-Native Aspect Ratios

Choose 16:9, 4:3, 1:1, 3:4, or 9:16. Output is rendered at your target ratio natively — not cropped from a wider shot. Match your input image ratio to the selected size to avoid distortion in the Cosmos 3 Super image-to-video output.

Commercial Use on Every Plan

Every clip generated here is covered by the NVIDIA OpenMDW-1.1 license. Use the output in paid social ads, product videos, client deliverables, and platform-distributed content. Attribution required: "Built on NVIDIA Cosmos." No tier-gated commercial-use unlock.

Who Uses the Cosmos 3 Super AI Video Generator

Real-world workflows where Cosmos3-Super-Image2Video delivers results faster than traditional video production.

01 / E-commerce & Product

Turn Product Photos into Video Ads

Upload a product still — a shoe, a perfume bottle, a jacket — and animate it with a motion prompt: rotating camera, liquid pour, fabric movement. The physics-real output looks like studio footage, not AI animation. Generate multiple variants by changing the motion prompt and re-running.

02 / Social Media Content

Animate Static Assets for Reels and TikTok

Select 9:16 output and animate any portrait image, illustration, or brand graphic into a vertical clip. Cosmos3-Super-Image2Video preserves visual identity while adding motion — no footage library, no video editor, no stock license required. Generate, download, post.

03 / Advertising & Agencies

Produce Multiple Creative Variants Fast

Agencies can generate multiple motion treatments from a single hero image by changing the motion prompt and duration. Every output is commercially licensed under OpenMDW-1.1. A/B test creatives for Meta or Google campaigns without a production team or video shoot.

04 / Pre-visualization & Concept

Animate Storyboard Frames Before Shooting

Feed a concept illustration or pre-vis still into the Cosmos 3 Super image-to-video model and generate a rough motion test. Directors use it to validate camera behavior and lighting intent before committing to a real shoot — cutting rounds of revision out of the production cycle.

Cosmos 3 Super AI Video Generator vs Alternatives

How Cosmos3-Super-Image2Video compares on the dimensions creators and production teams actually care about.

DimensionCosmos 3 Super I2VKling 3 I2VWan 2.1 I2VRunway Gen-4 I2VLuma Ray 3 I2V
Open Weights Yes No Yes No No
Physics-IQ Rank#1 open models
Ambient Audio (same pass) Built-in
Commercial Use (all plans) OpenMDW-1.1Paid tier only Apache 2.0Paid tier onlyPaid tier only
No GPU / Browser-Based On this siteSelf-host only
Negative Prompt Support Native parameterLimitedLimited
Max Duration7 seconds10 seconds5 seconds10 seconds9 seconds
Aspect Ratios16:9 · 9:16 · 1:1 · 4:3 · 3:416:9 · 9:16 · 1:116:9 · 9:1616:9 · 9:16 · 1:116:9 · 9:16 · 1:1

Cosmos 3 Super AI Video Generator — FAQ

The Cosmos 3 Super AI Video Generator — powered by Cosmos3-Super-Image2Video — is a post-trained specialist model from NVIDIA's Cosmos 3 family. Released May 31, 2026, it animates a single reference image into a physics-real video clip using a text-based motion prompt. It is built on a 64B Mixture-of-Transformers architecture and ranks #1 on the open Physics-IQ benchmark.
No. This site runs Cosmos3-Super-Image2Video entirely on cloud infrastructure. You upload your image and submit a prompt from any browser — no local GPU, no Python environment, no CUDA setup required on your end. The underlying model runs on NVIDIA Hopper and Blackwell datacenter GPUs on the server side.
Yes. The Cosmos 3 model family is released under the NVIDIA OpenMDW-1.1 license, which permits commercial use including paid advertisements, client deliverables, and content distributed on platforms. All plans on this site include commercial use. You are required to include a "Built on NVIDIA Cosmos" attribution per the license terms.
Cosmos3-Super-Image2Video supports 16:9, 4:3, 1:1, 3:4, and 9:16 output aspect ratios. For best results, match your input image aspect ratio to the selected output size. If the ratios don't match, the model may stretch or distort the output to fit the target dimensions.
Cosmos3-Super-Image2Video is currently the only open-weights image-to-video model ranked #1 on the Physics-IQ benchmark — it simulates real-world motion (fluid dynamics, fabric drape, object collisions) more accurately than alternatives. It also generates synchronized ambient audio in the same forward pass, which no current competitor does natively. Commercial use is included on every plan without a tier upgrade.
Cosmos3-Super-Image2Video accepts JPG, PNG, JPEG, and WEBP files in RGB color (8 bits per channel, sRGB color space). Grayscale images are not supported. Supported input resolutions are 256p, 480p, and 720p. Maximum file size on this site is 10 MB.
No. Your uploaded images and motion prompts are used solely to run your generation request. They are not used to retrain, fine-tune, or improve any model. See our Privacy Policy for full data handling details.

Animate Your Images with Cosmos 3 Super

Physics-real video from a single still image — in your browser, on any plan.