NVIDIA Cosmos 3 Super
Write a prompt, pick one of five aspect ratios, and NVIDIA Cosmos 3 Super turns it into an image for 8 credits. A finished frame can go straight into the image-to-video variant and become a clip of up to 7 seconds.
A selection of published prompt entries. Every card shows the prompt text — open one, read it, then write your own in the box above.
This painterly landscape features towering cumulus clouds mirrored in a calm river, set against patchwork farmland and a village church. The romantic fine-art style emphasizes serene pastoral elegance. Suitable for art exhibition displays and nature-themed illustration projects.
This image presents a Venetian canal in an oil painting style, featuring rich blue and gold tones with dramatic lighting. It captures a romantic, dreamlike travel poster atmosphere. Suitable for tourism brochures and city landscape illustrations.
A butterfly rests within a golden wheat field, with soft light and delicate textures lending the image a romantic, artistic atmosphere. Its natural subject and warm setting create a gentle visual focus. Suitable for nature illustrations and romantic poster designs.
This image captures a golden retriever resting by a campfire at a mountain lake, featuring warm firelight and serene wilderness details. Suitable for animal theme photography and outdoor landscape illustration projects.
This image presents a dreamlike forest avenue featuring symmetrical trees and a colorful geometric ground, evoking a stained-glass aesthetic with soft, layered lighting. Suitable for artistic illustration projects and game environment design concepts.
Toronto waterfront glows under golden hour light, where falling red leaves frame the city skyline. This serene autumn scene evokes nostalgic romance through warm tones and tranquil atmosphere. Suitable for travel photography projects and seasonal graphic design.
This warm portrait captures a smiling young South Asian woman in a cozy, plant-filled interior. Soft lighting and shallow depth of field create an inviting, realistic atmosphere. Suitable for character portrait posters and lifestyle magazine covers.
An older florist holds a bunch of orange dahlias at dusk, photographed in warm grainy light inside a flower shop. Suitable for editorial portraits and lifestyle illustration projects.
These images come from our prompt library and were not all made with Cosmos 3 Super. Your results will look different.
Cosmos 3 Super is the largest variant of NVIDIA's Cosmos 3 model family, released on May 31, 2026 on Hugging Face and GitHub. The family comes in three sizes: Super at 64 billion parameters, Nano at 16B, and Edge at 4B. On this site, Super powers two variants — Cosmos3-Super-Text2Image for still images and Cosmos3-Super-Image2Video for clips built from a start image.
The site gives the models a web workspace. You pick an aspect ratio, write a prompt, and pay per generation in credits: 8 for each image and 10 per second of video. The same dashboard also hosts other image and video models behind one balance, so you can compare outputs without switching tools.
Cosmos 3 AI is an independent site and is not affiliated with, sponsored by, or endorsed by NVIDIA.
| Developer | NVIDIA |
|---|---|
| Model | Cosmos 3 Super (~64B parameters) |
| Released | May 31, 2026 — Hugging Face & GitHub |
| Input | Text prompt, or one image + prompt for video |
| Output | JPEG image (PNG selectable) or MP4 clip |
| Aspect ratios | 1:1, 3:4, 9:16, 4:3, 16:9 |
| Video length | 1–7 seconds |
| Cost | 8 credits per image · 10 credits per second of video |
| Free credits | 10 for new accounts |
| Runs on | This site — nothing to install |
The hero box hands your prompt and settings to the dashboard — nothing is generated until you press the button there.
Type what you want in plain words in the box above — a subject, a setting, and a style are usually enough for a solid first result.
Open the aspect panel and choose one of the five ratios. 1:1 suits profile pictures and posts; 16:9 fits banners and wide scenes; 9:16 is built for vertical feeds.
In the dashboard you can add a negative prompt and adjust inference steps (1–50, default 28) and guidance scale (0–20, default 4) before you generate.
Each image costs 8 credits and lands in your history. Download it, or send the frame to the image-to-video workspace to animate it.
Six jobs the text-to-image variant handles well.
Lighting, reflections, and material texture hold together in complex scenes — mist over water, firelight on fur, metal on a workbench. Prompts that name the light source and the surfaces get the most out of the model.
Clean studio-style renders of bottles, packaging, and gadgets with controlled lighting and backgrounds. Useful for concept ads and mock campaigns before a real shoot.
Rooms, facades, and furniture with consistent perspective and believable daylight. Works for mood boards, staging ideas, and environment concepts.
Wide establishing shots with atmosphere — weather, depth, and a strong focal point. Useful for storyboards, game environments, and keyframe sketches.
The 9:16 and 1:1 ratios produce feed-ready compositions without cropping afterward — portraits, products, and scenes framed for the platform they land on.
Any image you generate can become the first frame of a clip. Send it to the image-to-video workspace, describe the motion, and get an MP4 back.
Open the image-to-video pageThese images come from our prompt library and were not all made with Cosmos 3 Super. Your results will look different.
The same controls appear in the hero box and in the full dashboard.
Tell the model what to leave out — text, blur, artifacts — using the negative prompt field in the dashboard's optional settings.
Tune inference steps (1–50, default 28) and guidance scale (0–20, default 4) per generation instead of accepting fixed defaults.
1:1, 3:4, 9:16, 4:3, and 16:9 are first-class settings for both modes, not crops applied afterward.
A finished frame moves into the video workspace in one step and comes back as a 1–7 second MP4 clip.
The dashboard hosts more image and video models that draw from the same credit wallet, so you can compare outputs without switching tools.
The marketing pages, dashboard, and legal pages are all localized — pick your language in the header.
The two Cosmos 3 Super variants side by side, as the dashboard exposes them.
| Parameter | Text to Image | Image to Video |
|---|---|---|
| Input | Text prompt | One start image + text prompt |
| Output | JPEG image (PNG selectable) | MP4 video, no audio |
| Aspect ratios | 1:1, 3:4, 9:16, 4:3, 16:9 | Same five — start image is center-cropped |
| Prompt length | Up to 2,500 characters | – |
| Inference steps | 1–50, default 28 | 1–50 |
| Guidance scale | 0–20, default 4 | 0–20, default 6 |
| Duration | – | 1–7 seconds |
| Cost | 8 credits per image | 10 credits per second |
Every model in the dashboard draws from the same credit balance. This table lists the cost per generated image for a few of them.
| Model | Developer | Cost per image | Good fit for |
|---|---|---|---|
| Cosmos 3 Super | NVIDIA | 8 credits | Images plus image-to-video in one model |
| Z Image Turbo | Alibaba | 2 credits | High-volume drafts at the lowest cost |
| NanoBanana 2 Lite | 6 credits | Editing existing images | |
| WAN 2.7 | Alibaba | 8 credits | Photoreal subjects |
| GPT Image 2.0 | OpenAI | 10 credits | Detailed instruction following |
Generation is billed in credits: 8 for each Cosmos 3 Super image and 10 per second of video. Top up with a credit pack whenever you need more, or subscribe for a monthly allowance. Credit-pack credits never expire.
See pricingA roughly 64-billion-parameter model from NVIDIA's Cosmos 3 family, released on May 31, 2026. On this site it runs two variants: Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video.
New accounts receive 10 free credits on first claim — enough for one image or a 1-second clip. After that, images cost 8 credits each and video costs 10 credits per second.
The weights are open on Hugging Face, but Super needs multi-GPU H100-class servers to serve — Hugging Face recommends an 8×H100 node. On this site it runs in the browser; there is nothing to install.
1:1, 3:4, 9:16, 4:3, and 16:9, for both image and video. For video, the start image is resized and center-cropped to the ratio you choose.
Yes — the image-to-video variant animates one start image plus a prompt into a 1–7 second MP4 clip. There is no text-to-video mode.
Yes. The dashboard exposes a negative prompt field in the optional settings, for both image and video generation.
The credits reserved for that task are refunded to your balance automatically.
Outputs made on a paid plan may be used commercially under our Terms. If you self-host the model instead, check NVIDIA's model license. Read the Terms
No. Cosmos 3 AI is an independent site that hosts the open Cosmos 3 Super weights through a third-party inference API. It is not affiliated with, sponsored by, or endorsed by NVIDIA.
Yes. The dashboard hosts additional image and video models that draw from the same credit balance, so you can compare results without a second account.
An account. Signing in opens the dashboard and 10 free credits on first claim.
Write a prompt in the box above and generate your first image.
