ComfyUI AWS Integration Nodes

This repository contains custom nodes for ComfyUI that integrate with AWS services for AI image generation:

1. Bedrock Node: Image-to-image generation using Amazon Bedrock and Stable Diffusion

2. SageMaker Node: Text-to-image generation using FP8 Flux Dev 1 diffusion pipeline

Both nodes seamlessly integrate with your ComfyUI workflows, allowing you to leverage AWS's powerful AI infrastructure directly within the ComfyUI interface.

Features

Bedrock Node

• Image-to-image generation using Stability AI's SD3 model through Amazon Bedrock
• Automatic image resizing to meet model requirements
• Support for customizable prompt and strength parameters

SageMaker Node

• Text-to-image generation using FP8 Flux Dev 1 diffusion pipeline
• Custom FP8-optimized model deployment
• Flexible parameter customization

Prerequisites

• An AWS account with access to Amazon Bedrock and SageMaker
• Proper AWS credentials configured
• Python 3.x
• ComfyUI installed

Installation

1. Configure your AWS credentials through one of these methods:

   • AWS CLI (aws configure)
   • Environment variables
   • AWS credentials file

2. Add both node folders to the ComfyUI custom_nodes directory

Usage

Bedrock Node

1. Launch ComfyUI
2. Find the "Image2ImageNode" in the node browser under the "image" category
3. Connect the node to your workflow with:
   • Input image
   • Text prompt
   • Strength value (0-1)

Parameters

• image: Input image tensor
• prompt: Text description for the desired image modification
• strength: Float value between 0 and 1 determining how much to modify the original image (default: 0.75)

SageMaker Node

1. Deploy the FP8 Flux Dev 1 model using the provided notebook
2. Launch ComfyUI
3. Find the SageMaker "Text@Image" node in the node browser
4. Connect the node to your workflow with:
   • Text prompt
   • Optional parameters

Model Deployment

There was no full pipeline available that leveraged the flux1dev transformer in FP8, so I created one and made it available here: Jlonge4/flux-dev-fp8. This is important as using hugging face diffusers from_single_file download option will create symlinks, rendering your model.tar.gz unusable for deployment.

To deploy the model:

1. Run the deploy_flux_dev-pipe.ipynb notebook
2. A g5.8xlarge instance is recommended for deployment
3. Once deployed, the endpoint can be used directly by the SageMaker node

Notes

Bedrock Node

• Automatically handles image resizing to meet the model's requirements (minimum 640px, maximum 1536px)
• Images are processed maintaining aspect ratio
• Uses the stability.sd3-large-v1:0 model from Amazon Bedrock

SageMaker Node

• Be mindful of the 60 second timeout for SageMaker inference endpoints
• The single FP8 safetensors file I used to create my pipeline / HF repo is found here: Comfy-Org/flux1-dev.