Cosmos 3 Supported Video Resolutions and Frame Rates: Complete Generation Settings Guide

Cosmos 3 supports three fixed resolution tiers (256p at 320×192, 480p at 832×480, and 720p at 1280×720) and four frame rates (10, 16, 24, and 30 FPS), with 480p and 24 FPS as the default settings.

The NVIDIA Cosmos repository defines the supported video resolutions and frame rates in Cosmos 3 through a strict set of spatial and temporal constraints documented in the source code. These fixed parameters govern all generation modes, including text-to-video, image-to-video, and text-to-image outputs, ensuring consistent performance across the Cosmos 3 foundation model.

Supported Video Resolutions in Cosmos 3

Cosmos 3 organizes output quality into three discrete resolution tiers. Each tier maps to specific pixel dimensions, and these represent the only supported spatial configurations for any generation request.

256p Low Resolution (320×192)

The 256p tier generates content at 320 pixels wide by 192 pixels tall. This resolution is suitable for rapid prototyping and scenarios where generation speed takes priority over visual fidelity.

480p Standard Resolution (832×480)

The 480p tier outputs 832×480 pixels and serves as the default resolution for Cosmos 3 generation. This setting balances computational efficiency with output quality for standard video generation tasks.

720p High Resolution (1280×720)

The 720p tier produces 1280×720 pixels, offering the highest spatial fidelity available. This setting requires significantly more GPU memory and is ideal for final production-quality outputs.

Supported Frame Rates for Video Generation

Cosmos 3 supports four specific frame rates measured in frames per second (FPS): 10, 16, 24, and 30 FPS. The 24 FPS setting is the default value applied when no specific frame rate is requested.

These frame rates apply exclusively to video generation with multiple frames. For single-frame image generation, the fps parameter is accepted by the pipeline but ignored during processing.

Configuring Resolution and Frame Rate in Code

When using the Cosmos 3 Diffusers pipeline, you specify resolution through the height and width parameters and temporal settings through the fps parameter. The following example demonstrates configuring the Cosmos3OmniPipeline for different output specifications.

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

# Load the Cosmos 3 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

result_720p = 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,
    enable_sound=False,
)

# Generate 256p image (single frame)

result_256p = pipe(
    prompt="A tiny robot exploring a miniature garden.",
    num_frames=1,
    height=192,
    width=320,
    fps=24,  # Ignored for single-frame generation

    num_inference_steps=30,
    guidance_scale=5.0,
)

# Save outputs

pipe.save_output(result_720p, "video_720p_30fps.mp4")
pipe.save_output(result_256p, "image_256p.png")

CLI and Framework Configuration

According to the Cosmos source code, you can also specify these parameters through the Cosmos Framework CLI using form strings. As implemented in cookbooks/cosmos3/generator/action/run_policy_with_cosmos_framework.md, pass resolution and frame rate settings via command-line arguments:

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

The size parameter accepts the pixel dimensions directly (1280×720 for 720p, 832×480 for 480p, or 320×192 for 256p), while the fps parameter accepts one of the four supported integer values. Additional implementation examples appear in cookbooks/cosmos3/generator/audiovisual/README.md, which illustrates practical usage of these constraints in notebook environments.

Source Documentation Reference

The authoritative list of supported generation settings appears in the repository's top-level README.md (lines 89-94). This documentation explicitly defines the valid resolution tiers and frame rates, noting that selecting a resolution tier automatically determines the output dimensions and influences the maximum frame count for video generation to maintain tractable computational workloads.

Summary

• Cosmos 3 supports three resolution tiers: 256p (320×192), 480p (832×480), and 720p (1280×720), with 480p as the default.
• Four frame rates are available: 10, 16, 24, and 30 FPS, defaulting to 24 FPS for video generation.
• Resolution tiers are fixed: The model does not support arbitrary resolutions; you must use the specific pixel dimensions associated with each tier.
• Configuration occurs via pipeline parameters: Use height, width, and fps in the Diffusers pipeline or form strings in the CLI framework.
• Source documentation: Official settings are documented in README.md lines 89-94 and implemented throughout the cookbooks.

Frequently Asked Questions

What is the highest resolution supported by Cosmos 3?

The highest supported resolution is 720p at 1280×720 pixels. This corresponds to the 720p resolution tier. Cosmos 3 does not support 1080p or 4K resolutions; attempting to specify dimensions outside the three defined tiers will result in errors or undefined behavior.

Can I use custom frame rates like 60 FPS or 25 FPS?

No. Cosmos 3 strictly supports only 10, 16, 24, and 30 FPS. The model architecture and training data constrain temporal generation to these specific values. If you require different playback speeds, you must generate at one of the supported rates and perform frame rate conversion post-generation using external video processing tools.

How does resolution affect the maximum video length?

Higher resolution tiers impose stricter limits on the maximum number of frames to maintain computational tractability. While the default generation uses 189 frames (yielding approximately 7.9 seconds at 24 FPS), the resolution tier you select influences the practical upper bound for video length. The 720p tier requires significantly more memory than 256p, effectively limiting the maximum duration for high-resolution outputs on standard GPU configurations.

Where are the supported generation settings defined in the source code?

The definitive list appears in the top-level README.md at lines 89-94 under the "Supported Generation Settings" section. This table documents the resolution tiers (256p, 480p, 720p) and frame rates (10, 16, 24, 30 FPS). Implementation examples reside in cookbooks/cosmos3/generator/audiovisual/README.md and cookbooks/cosmos3/generator/action/run_policy_with_cosmos_framework.md, which demonstrate how these constraints propagate through the Diffusers pipeline and CLI framework respectively.

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