Stability Creative Upscaler

The creative member of this pack's upscaler trio is the one that's actually allowed to make things up. Where the conservative node is locked to a tight 0.2–0.5 creativity band and told to behave, StabilityUpscaleCreative gets a dial that opens lower and a mandate to add detail that was never in the source. It's Stability's hosted /v2beta/stable-image/upscale/creative endpoint, and it's the SUPIR-class job in this pack: rebuild the image at higher resolution with invented-but-plausible texture.

Reach for it when the source is soft, small, or under-detailed and you want the model to construct the missing information - think printing an old render, or adding fabric weave and pores to a flat image. Don't reach for it when identity matters. Every generative upscaler rewrites faces and fine structure at high creativity; the community's standing rule is that a face you need to keep recognizable gets its own pass at low strength. This node is the same deal, at cloud prices.

How it works - the different part

Unlike the fast and conservative nodes, this one is asynchronous. It POSTs your image and waits, gets back a generation ID, then polls Stability's /v2beta/results/{id} endpoint every 10 seconds until the upscale is done - up to a hard 5-minute timeout. What that means in ComfyUI: the node blocks the queue while it waits, so a workflow that chains several of these can feel sluggish even though the API is doing the real work.

The prompt matters even more than on the conservative node. Creative upscaling is generation, and the model needs an accurate caption to know what detail to add. Vague prompts produce confident but wrong detail.

Inputs and outputs

• image, prompt - both required.
• creativity - 0.1–0.5, default 0.3. Treat this as a leash, not a slider to max: above ~0.4 the output stops being an upscale of your image and starts being a different picture wearing its clothes.
• negative_prompt, seed - standard.
• style_preset - 18 named looks; none unless you want the upscale restyled.
• output_format - png / jpeg / webp.
• api_key - per-node override.

Output: a single IMAGE, into Save Image as usual.

Installing it

Part of the aicu-comfyui-stability-ai-api pack:

cd ComfyUI/custom_nodes
git clone https://github.com/aicuai/aicu-comfyui-stability-ai-api.git
cd aicu-comfyui-stability-ai-api
pip install -r requirements.txt

Or install via ComfyUI Manager and restart. You need a paid Stability AI key from platform.stability.ai - node field, STABILITY_API_KEY env var, or the pack's auto-generated config.ini.

Gotchas

Creative upscaling is the most expensive call in the trio, so budget before you batch fifty images. Also note the input cap: total pixels must stay under ~1MP (4,096 to 1,048,576), which is smaller than the conservative node's ~9.4MP allowance - feed a 2048px image in raw and it'll error before you spend anything, so downscale first. And the 5-minute timeout is real: a long queue on Stability's side and the node throws "timed out after 5 minutes." Just rerun it - you only pay for completed generations. For "bigger, sharper, and I don't mind if it adds a little character," this is the one.