Cosmos 3 Supported Resolutions and Frame Rates: Complete Technical Guide

Cosmos 3 supports three resolution tiers (256p, 480p, and 720p) and four frame rates (10, 16, 24, and 30 FPS), with 480p and 24 FPS set as defaults for all generation tasks.

NVIDIA Cosmos 3 is an open-source generative AI model designed for text-to-image, text-to-video, and image-to-video synthesis. According to the NVIDIA Cosmos repository, the model architecture enforces strict constraints on output dimensions and temporal frequency to ensure computational tractability and generation quality.

Resolution Tiers and Pixel Dimensions

Cosmos 3 defines three fixed resolution tiers that determine the spatial dimensions of generated outputs. These values are hardcoded in the model configuration and documented in the repository’s top-level README.md (lines 89-94).

The supported tiers map to the following width × height dimensions:

• 256p → 320 × 192 pixels
• 480p → 832 × 480 pixels (default)
• 720p → 1280 × 720 pixels

Selecting a resolution tier automatically constrains the output dimensions. For video generation, the chosen tier also influences the maximum frame count to maintain memory efficiency during inference.

Frame Rate Options

For video outputs, Cosmos 3 supports four discrete frame rates that control playback speed:

• 10 FPS
• 16 FPS
• 24 FPS (default)
• 30 FPS

The frame rate parameter determines the temporal resolution of the generated MP4 file. For example, when using the default 189-frame length, 24 FPS produces approximately 7.9 seconds of video, while 30 FPS yields roughly 6.3 seconds. According to the source code in README.md, these are the only supported temporal frequencies for any Cosmos 3 video generation request.

Implementation in Code

You specify resolutions and frame rates via the height, width, and fps parameters in the Cosmos3OmniPipeline. The following examples demonstrate valid configurations using the Diffusers library:

import torch
from diffusers import Cosmos3OmniPipeline
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler

# Load the Nano checkpoint

pipe = Cosmos3OmniPipeline.from_pretrained(
    "nvidia/Cosmos3-Nano",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)

# Generate 720p video at 30 FPS (1280×720)

video_result = pipe(
    prompt="A sleek autonomous vehicle drives through a futuristic city at night.",
    num_frames=189,
    height=720,
    width=1280,
    fps=30,
    num_inference_steps=35,
    guidance_scale=6.0,
)

# Generate 256p image at 24 FPS (320×192, fps ignored for single frames)

image_result = pipe(
    prompt="A tiny robot exploring a miniature garden.",
    num_frames=1,
    height=192,
    width=320,
    fps=24,
    num_inference_steps=30,
)

pipe.save_output(video_result, "output_720p_30fps.mp4")
pipe.save_output(image_result, "output_256p.png")

CLI and API Usage

When using the Cosmos Framework CLI or the vLLM-Omni REST API, you pass resolution and frame-rate values through form strings or JSON parameters. As documented in cookbooks/cosmos3/generator/action/run_policy_with_cosmos_framework.md, the CLI accepts:

--form-string "size=1280x720" \
--form-string "fps=24"

The API validates these inputs against the supported tiers listed in cookbooks/cosmos3/generator/audiovisual/README.md. Requests specifying unsupported dimensions or frame rates will return validation errors before inference begins.

Summary

• Cosmos 3 enforces three fixed resolution tiers: 256p (320×192), 480p (832×480), and 720p (1280×720).
• Four frame rates are supported: 10, 16, 24, and 30 FPS, with 24 FPS as the default.
• The 480p tier and 24 FPS are used automatically when parameters are omitted.
• Resolution tiers affect both image dimensions and maximum video length constraints.
• All valid combinations are documented in the repository’s README.md and enforced by the Cosmos3OmniPipeline class.

Frequently Asked Questions

What is the default resolution and frame rate for Cosmos 3 generation?

The default resolution is 480p (832×480 pixels) and the default frame rate is 24 FPS. These values are applied automatically when you omit the height, width, or fps parameters in API calls or CLI commands.

Can I use custom resolutions outside the three supported tiers?

No. Cosmos 3 does not support arbitrary resolutions. The model architecture requires inputs to conform to one of the three predefined tiers (256p, 480p, or 720p). Attempting to specify dimensions like 1024×576 will result in validation errors or undefined behavior.

How does frame rate selection affect video generation time?

Higher frame rates do not increase generation time, but they do reduce the effective duration of the output video. For example, with the standard 189-frame limit, 30 FPS produces a 6.3-second video, while 10 FPS yields 18.9 seconds of footage. The inference cost remains constant regardless of the FPS setting.

Are the resolution and frame rate settings used for image generation?

Resolution tiers apply to both images and videos. For single-image generation (num_frames=1), you still specify height and width according to the tier dimensions. However, the fps parameter is ignored for image outputs since static images have no temporal component.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}