Cosmos3-Nano — Video Generation AI Model
Cosmos3 Nano is NVIDIA's 16B world model, built to make video that obeys physics rather than merely looking plausible. Text in, an image to anchor frame zero, or a clip to continue. Optional synced audio — describe the sound in the prompt.
Demo
Output from the prompt “a close macro shot of morning dew sliding down a green fern frond, soft daylight, shallow depth of field”
What it does
Cosmos3 Nano generates video from a written description, animates a still first frame, or picks up where a clip you supply ends. It is NVIDIA's 16B world model, built to produce motion that obeys physics — things fall, water runs and light moves the way they do rather than merely looking plausible frame by frame.
One run yields one clip of about eight seconds at 1280x720. The prompt sets both the look and the motion; with an input image the look is anchored to that frame and the text drives only what happens. Audio is off by default — describe the sound in the prompt to turn it on.
When to use it
- Short establishing shots and B-roll from a written description
- Animating a still image while keeping its exact look in frame zero
- Previewing a scene before committing to a live shoot
- Motion studies where physical plausibility matters more than styling
- Extending a clip past one run by feeding each run the previous one
Details
- CategoryVideo Generation
- Year2026
- LicenseOpenMDW-1.1
Compliance & Provenance
- ProviderNVIDIA (open) · Specialized
- EU AI Act RiskLimited Risk
- Art. 50 TransparencyRequired — AI-generated outputs are marked
Inputs & Outputs
- TextInputstring
Describe the scene and motion in one plain paragraph. With an input image, frame 0 carries the look and this text drives what happens over time; with an input video, the clip picks up where that video ends and this text drives what happens next; without either, the text alone sets both look and motion (text-to-video). When sound is on, mention the audio you want too.
- ImageInputimageoptional
Optional first frame. When provided, frame 0 is anchored to it (image-to-video); when omitted, the whole clip is generated from the text prompt alone (text-to-video).
- VideoInputvideooptional
Optional clip to continue (at least 5 frames). When provided, the new clip starts from this clip's last 5 frames and carries its motion on; the image input is then ignored. The first 5 frames of the output repeat the end of this clip - drop them when joining the two. Best with a 24 fps clip at the chosen resolution's aspect ratio: it is stretched to that resolution.
- VideoOutputvideo
Generated video
Parameters
- resolutionstringdefault 1280x720 (16:9)options: 1280x720 (16:9), 720x1280 (9:16), 960x960 (1:1), +7 more
Resolution (W×H) — pick the ratio matching the input image or video
Output resolution. Values follow the official cosmos-framework resolution buckets: 720p tier — 1280x720 (16:9), 720x1280 (9:16), 960x960 (1:1), 1104x832 (4:3), 832x1104 (3:4); 480p tier — 832x480, 480x832, 640x640, 736x544, 544x736. Buckets within a tier keep roughly equal pixel area. Default '1280x720 (16:9)' matches the official Cosmos3-Nano examples and the pipeline default. The 480p tier roughly halves VRAM and latency vs 720p. Pick the ratio closest to the input image or video — either is stretched to the requested resolution, so a mismatched ratio distorts the start of the clip.
- num_framesstringdefault 189 frames (~7.9s @ 24fps)options: 61 frames (~2.5s @ 24fps), 121 frames (~5s @ 24fps), +3 more
Frame count (4n+1 values)
Total frames to generate. The official supported range is 5-400 frames; 189 (default, ~7.9s at 24 fps) is the official default used in all examples. Presets are chosen on the 4n+1 grid of the VAE temporal compression (not officially enforced — the pipeline floors other values internally): '61 frames (~2.5s)', '121 frames (~5s)', '189 frames (~7.9s)' (default), '241 frames (~10s)', '301 frames (~12.5s)'. Longer clips are proportionally more expensive in VRAM/time.
- fpsstringdefault 24options: 16, 24, 30
Output FPS (24 native)
Target frame rate. Unlike container-only fps settings, Cosmos3 injects fps into the temporal RoPE modulation and duration metadata, so it conditions the generated motion. 24 (default) is the training-native rate used in all official examples — recommended. 16 or 30 shift the motion speed/duration trade-off.
- enable_soundstringdefault Offoptions: On, Off
Sound generation (On/Off) — describe the desired sound in the prompt
When 'On', Cosmos3-Nano jointly generates a synchronized audio track (48kHz, muxed into the mp4 as AAC) using its built-in sound tokenizer, at the cost of some extra latency and VRAM. IMPORTANT: when 'On', describe the desired audio in the text prompt (e.g., 'quiet room ambience with soft rustling; no speech, no narration, no music') — without an explicit audio description the model invents arbitrary sound, often unintelligible speech-like narration. Default 'Off' — the official image-to-video example generates silent video.
- stepsintegerdefault 35range 20–75
Diffusion steps (20-75)
Number of denoising steps. Default 35 matches the Cosmos3OmniPipeline default used by the official Cosmos3-Nano examples. Raise toward 50-75 if visible artifacts remain; drop toward 20 to save time at the cost of quality.
- guidance_scalefloatdefault 6range 1–10advanced
CFG scale
Classifier-free guidance scale. Default 6.0 — official Cosmos3 recommendation. Higher pushes the output toward the prompt at the cost of motion diversity; lower allows freer composition.
- negative_promptstringdefault advanced
Additional negative prompt
Extra negative prompt text appended to Cosmos3's recommended quality-control negative prompt (which already covers compression artifacts, flicker, implausible physics, etc.). If the Text input is an upsampled JSON prompt containing its own negative_prompt, that one is used as the base instead. Leave empty for the default.
- seedintegerdefault -1range -1–2147483647advanced
Random seed (-1 = random)
Random seed for reproducibility. -1 (default) selects a fresh random seed every run. Set a fixed non-negative integer to reproduce a previous video (same seed + same image + same prompt + same params -> same output).
Limitations
- One clip per run, around eight seconds. For longer sequences, feed each run's output into the next run's Video input: the motion carries across the join, but the subject's look can still drift over many runs.
- The prompt describes, it does not direct. Precise framing, camera moves and timing are not controllable the way they are in an editor.
Tags
Alternatives in Video Generation
- Helios Base
On-device text-to-video, up to 60s @ 24fps. 3-10 min.
- Helios Distilled
Fast distilled text-to-video, fixed 640x384. 1-5 min.
- LatentSync 1.6 (Lip Sync)
Re-sync a talking video's lips to a new voice.
- LTX-2.5
Lightricks LTX-2.5 distilled 22B: text, or text plus a first-frame image, to a video with a synchronized soundtrack generated alongside it.
- LTX-2.5 Audio-to-Video
Make a talking video from one picture and a voice.
- MiniMax-H3 Ref2VA
MiniMax-H3 Ref2VA (omni-reference) generates a 5-15s 24fps video with native stereo audio from a prompt plus a reference image and a reference clip, that clip's own soundtrack included. Video and its soundtrack come out of one denoising loop, so lip movement and sound land in sync. ~144GB bf16 weights — offloaded component-by-component, minutes-scale per clip.
- Wan2.2 TI2V 5B
Wan2.2 dense 5B unified text+image-to-video. Text only → 5s video from prompt; text + image → image as the first frame of the video.