Amazon Bedrock - Stability AI Models | Text to Image

Four Stability AI text-to-image models behind one dropdown, all running on AWS: SD3 Large, SD3.5 Large, Stable Image Core, and Stable Image Ultra. Type a prompt, pick a model, get an image - your GPU is irrelevant to the whole transaction.

How it works

The node maps your dropdown choice to a Bedrock model ID - stability.sd3-large-v1:0, stability.sd3-5-large-v1:0, stability.stable-image-core-v1:1, or stability.stable-image-ultra-v1:1 - and calls the Stability endpoint in us-west-2. It converts your chosen resolution into an aspect ratio (that's why "1024 x 1024" becomes 1:1, "1536 x 640" becomes 21:9, and so on), then hands the prompt to the model server-side. The PNG that comes back is turned into a ComfyUI IMAGE tensor.

Worth knowing what you're actually picking. SD3 and SD3.5 are the open-weight models - SD3.5 is runnable locally, and the KB's stability history covers how the whole family shipped after the original SD team's departures. Stable Image Core and Ultra are the closed, API-first flagships: quality and text rendering on another level, but no weights, no local equivalent, and content policy attached. In this one node you get the full spread from "open model behind a managed API" to "pure proprietary API."

The inputs that matter

• model - the dropdown above (defaults to Stable_Image_Core). Core/Ultra are the better-looking ones for scenes; SD3.5 is the familiar open model if you want behavior you know.
• prompt - the whole ballgame. These are instruction-following models; structured natural-language prompts beat tag soup (see the KB's prompt-engineering essay - on this generation of models your prompt is a message, not a token bag).
• resolution - 9 aspect-ratio presets, not free-form pixels. The model picks exact dimensions within the ratio you choose.
• seed - pin it for repeatable outputs.
• negative_prompt (optional) - supported here, though it does less than it did in the SD1.5 era.

Output is a single IMAGE tensor (the source returns only the first generated image).

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 install via ComfyUI Manager (search "Amazon Bedrock"). Dependencies are light - boto3, requests, retry - because there are no model files to download.

The AWS side: enable model access for the Stability models you want in the us-west-2 region (this node hardcodes it), and have credentials configured (aws configure, or an IAM role on an AWS instance - the README's CloudFormation deploys a ready-made ComfyUI on SageMaker).

Where people get burned

• Region, again. us-west-2 for Stability vs us-east-1 for the pack's Nova Canvas nodes. Mixing families means enabling access in both regions.
• Model access is per-model. You must grant each of the four checkboxes you plan to use.
• The resolution list is ratios, not sizes. Don't expect "1536 x 640" to be honored to the pixel - it's the 21:9 request.
• Stable Image Ultra is strict. The closed models enforce content policy hard; a borderline prompt just gets refused.
• Billing. Per-image API calls, and Ultra/ Core are the pricier tier. For heavy experimentation, SD3.5 locally is cheaper than this node.

If your machine runs SD3.5 already, this node's only value is the Core/Ultra models you can't run at home. If you're GPU-less and want Stability quality in a ComfyUI graph, it's a solid shortcut - just mind the region.