NVIDIA Cosmos: text2video

Introduction

Cosmos is an advanced open - source model that has made a significant impact in the realm of AIGC, especially in the area of generating high - quality images from text prompts (text - to - image) and transforming existing images into new ones (image - to - image). It offers a unique set of capabilities that allow for creative and realistic image generation.

Workflow Overview

Installation of Nodes and Models

Nodes:

• For ComfyUI, which is a popular framework to use with Cosmos, you need to ensure that it is up - to - date. Some Cosmos - specific nodes might be available as custom extensions. These can usually be installed by following the official ComfyUI documentation on adding custom nodes. This often involves downloading the relevant node files and placing them in the appropriate ComfyUI directories.

Models:

• Text Encoder and VAE
  These can be sourced from dedicated repositories. For example, the text encoder oldt5_xxl_fp8_e4m3fn_scaled.safetensors and the VAE cosmos_cv8x8x8_1.0.safetensors are available at https://huggingface.co/comfyanonymous/cosmos_1.0_text_encoder_and_VAE_ComfyUI/tree/main. Place the text encoder in the ComfyUI/models/text_encoders directory and the VAE in the ComfyUI/models/vae directory.
• Diffusion Models
  The diffusion models, which are pivotal for the video generation process, can be found in safetensors format at https://huggingface.co/mcmonkey/cosmos-1.0/tree/main.
• They should be placed in the ComfyUI/models/diffusion_models folder. If you prefer the original .pt format, the official links can be found on the Hugging Face repositories, such as those related to text - to - video models like https://huggingface.co/nvidia/Cosmos-1.0-Diffusion-7B-Text2World