StabilityAI Suite - Creative Upscale

If you've ever stared at a 4x ESRGAN upscale and thought "bigger, but not better," you've hit exactly the problem this node exists to solve. Creative Upscale doesn't just resize - it gets a generation model to invent new plausible detail while it upscales, the Magnific-style "add detail" look. The catch: that model runs on Stability AI's servers, and every call burns API credits.

When would you reach for it? Local creative upscaling is a whole workflow - tile the image, run each tile through a ControlNet Tile pass, stitch it back, fight seams. This node collapses all of that into one node and one API call. No VRAM, no model downloads, no tiling math. If you want the "add detail" look without building the pipeline, it's the fastest path by miles.

How it works

The node posts your image and prompt to the stable-image/upscale/creative endpoint, gets back a job ID, then polls the result endpoint every 10 seconds until the image is ready (or gives up after about four minutes). Two things follow from that. First, it's async - expect to wait anywhere from half a minute to a couple of minutes depending on queue and image size. Second, that job ID gets written to log.txt in the pack's folder, because Stability only archives finished images for 24 hours. If your PC dies mid-poll, that ID is your ticket to getting the image back - see the Creative Upscale Recover File node in the same pack.

Inputs that matter

Most inputs you can safely leave alone, but three are worth your attention:

• image - the IMAGE you're upscaling.
• prompt - describe the content. It steers what detail gets hallucinated.
• creativity - the master dial, 0.01 to 0.35, default 0.3. Higher means more invented detail; past a point you get drift from the source. Start at the default and back off if the result stops looking like your image.

negative_prompt, seed, and output_format (png/webp/jpeg) are for reproducibility and file type. The output is a single IMAGE - wire it straight into a Save Image node.

Install

Standard custom-node fare. ComfyUI Manager: search "ComfyUI StabilityAI Suite" and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/florestefano1975/ComfyUI-StabilityAI-Suite

Restart ComfyUI. No requirements to pip-install - the pack only uses requests/Pillow/numpy/torch, which ComfyUI already ships. There's no model download either, because there's no local model.

Then give it an API key, one of three ways:

# 1. environment variable (recommended)
export SAI_API_KEY="sk-your-stability-key"

# 2. drop the key into sai_platform_key.txt inside the pack folder

# 3. per-node: paste it into the api_key_override field

Where people get burned

• It costs money. Every upscale bills against your Stability API credits. A single big upscale is fine; a batch of fifty shows up on your balance. Read the pricing before building a pipeline on it.
• Unauthorized means the key is missing or mistyped - check env, file, and override, in that order.
• Not enough credits is Stability telling you to top up.
• Crash mid-run - grab the ID from log.txt and use the Recover File node before the 24-hour window closes.

One thing worth knowing: florestefano1975 also maintains the popular Portrait Master and Prompt Composer packs, and this suite is a fork of Stability's own official nodes with extras bolted on. The pack itself has almost no community footprint on Reddit - you're mostly on your own for the API side, but the mechanics here are exactly what the code does.