Amazon Bedrock - SD3 & SD3.5 Large | Image to Image

SD3 and SD3.5 Large are the full-size Stability Diffusion models - the ones that want more VRAM than most people own. You can run a distilled version locally, or you can point this node at Bedrock and let AWS host the real thing for you. It does image-to-image: a starting image plus a prompt, remixed server-side.

How it works

The node takes your IMAGE, base64-encodes it, and calls Bedrock's Stability endpoint in the us-west-2 region with mode: "image-to-image". SD3.5 runs a full diffusion pass server-side - no local checkpoint, no GPU on your end. The response comes back as a PNG and gets converted back into a normal ComfyUI IMAGE tensor.

The KB's stability history is worth knowing here: SD3.5 is open-weight and runs locally if you have the hardware, so this node isn't a gateway to a model you can't have - it's the "I want the full-size model but won't buy the GPU" path, plus the convenience of not managing checkpoints at all.

The inputs that matter

• image - your starting point. Feed it from Load Image or Image From URL.
• model - SD3_Large or SD3.5_Large (default SD3.5, which is the better model - keep it).
• prompt - where you're steering it.
• strength (optional, default 0.6, 0–1) - the img2img denoise dial. Below ~0.5 you're doing a light remix that keeps composition; above 0.8 you're nearly generating from scratch with the input as a suggestion. This is the input you'll tune the most.
• negative_prompt (optional) - SD3 family still respects it.
• seed - pin it for reproducible results.

Output is a single IMAGE tensor. One honest limitation in the source: it returns only the first generated image, so this is a one-at-a-time node.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/aws-samples/comfyui-llm-node-for-amazon-bedrock.git
pip install -r comfyui-llm-node-for-amazon-bedrock/requirements.txt

Restart ComfyUI, or find it in ComfyUI Manager under "Amazon Bedrock". Deps are just boto3, requests, retry - nothing to download models for.

AWS setup matters more than usual here because of region: enable SD3/SD3.5 model access in us-west-2, not the us-east-1 that the Nova Canvas nodes in this pack use. Get the regions crossed and you'll stare at a credentials or access error with no idea why. Configure credentials via aws configure or an IAM role, or use the README's SageMaker CloudFormation deployment.

Where people get burned

• Region mismatch is the #1 failure. This node hardcodes us-west-2; the pack's Nova nodes hardcode us-east-1. Enable both if you use both.
• Model access is per-model. Granting "Stability AI" in the console isn't enough - the SD3 and SD3.5 checkboxes are separate lines.
• No resolution control. The output size is inherited from your input image; there's no resolution field here. Crop or resize before feeding if you want a specific aspect.
• Billing. It's a per-image API call, so a long tuning session adds up. Dial in strength with single generations before committing.

Reach for it when you want the real SD3.5 Large's img2img behavior - richer detail than the distilled local variants - without the hardware. If your rig runs SD3.5 locally fine, you're paying for something you already have.