Amazon Bedrock - Nova Canvas Generate Image

This is the workhorse of the Bedrock pack: a text-to-image node backed by Amazon's Nova Canvas model, served entirely over AWS. Type a prompt, get a picture - no checkpoint to download, no VRAM budget, no sampler to argue with. Your GPU sits this one out.

How it works

The node calls Bedrock's invoke_model against amazon.nova-canvas-v1:0 and hands over your prompt. Nova Canvas generates the image server-side and returns it as base64, which the node decodes into a standard ComfyUI IMAGE tensor. It's a fully managed API - the "closed-source model" end of the spectrum the KB's closed-source essay maps out, where the trade-off is convenience and quality against control, privacy, and per-image billing.

What makes it more than a dumb text-to-image box is the optional inputs, which switch the underlying API task:

• image + control_mode - give it a reference image and it does image-conditioned generation. control_mode is CANNY_EDGE or SEGMENTATION (canny line structure or segmentation map), steered by control_strength (default 0.7). That's the same idea as a local ControlNet, minus the preprocessor and the local model.
• color_palette - comma-separated hex codes like #FF0000,#00FF00,#0000FF switch the API to color-guided generation. Up to 10 colors, and the node validates them, so a typo raises a clear error instead of a weird image.

The inputs that matter

Required: prompt, num_images (1–5), resolution (28 presets, up to 4096 on the long side), cfg_scale (0–10, default 7), seed. Optional: image, negative_prompt, color_palette, control_mode, control_strength.

The two you'll actually touch: resolution (pick the output size, not an aspect-ratio hint - Nova Canvas renders at exactly that size) and seed (pin it for reproducibility; 0 means a random roll). negative_prompt works the way you'd expect, which is more than some modern local models can claim.

Installing and wiring 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 use ComfyUI Manager (search "Amazon Bedrock"). Light deps: boto3, requests, retry.

Then the AWS side, which is the part people skip: enable Nova Canvas model access in the us-east-1 region in the Bedrock console, and configure credentials (aws configure, or an IAM role if ComfyUI runs on an AWS instance - the README has a CloudFormation template that deploys the whole thing on SageMaker).

Where people get burned

• Region and model access. This node hardcodes us-east-1. If you enabled model access in another region, it errors. The Stability nodes in this pack hardcode us-west-2, so if you use both families you need access in both regions.
• Content moderation. Nova Canvas filters prompts; a flagged prompt fails with Content moderation blocked generation rather than producing output.
• Always premium. The pack hardcodes quality: "premium" in the generation config, so there's no cheap tier for experimentation. Nova Canvas is priced per image - a batch of 5 is five billable images.
• Conditioned generation demands its settings. If you feed an image, the node raises unless control_mode and control_strength are set. It won't silently guess.

Where it earns its place: when you want a hosted model's output inside a ComfyUI pipeline without buying the hardware, or when the rest of your stack is already AWS. If you're purely local, this node has nothing for you - that's fine, it's not aimed at you.