Nodes/Stability API nodes for ComfyUI/Stability Conservative Upscale
ComfyUI Node

Stability Conservative Upscale

Cloud upscaling that plays it straight — no invented detail

By Stability-AI·Created 2 years ago·Updated about a year ago· 64
Stability Conservative Upscale
  • image
  • IMAGE
prompt
negative_prompt
seed0
creativity0.35
output_format
api_key_override

The "conservative" in the name is the whole pitch. In upscaling you've got two jobs that get constantly confused: adding pixels (the image already has the detail you want, you just need it bigger) and adding detail (the model invents texture that was never there). Conservative Upscale is firmly in the first camp - it makes your image larger and keeps it faithful, rather than rewriting it into something glossier.

It's a cloud node from Stability's official "Stability API nodes for ComfyUI" pack, calling stable-image/upscale/conservative. It costs credits, needs a Stability API key, uses zero VRAM. The local alternative - a 4x ESRGAN model or Lanczos - is free, instant, and runs on a potato, and per the community's current upscaling mood that's often the better answer for the "more pixels" job. Where this node makes sense: you're on hardware that can't even run ESRGAN comfortably, or you want a generative-but-conservative upscale without babysitting a local model.

How it works - and the odd bit

Here's the wrinkle: even though it's the conservative upscaler, it still wants a prompt. That's not a bug - it's how the server stays faithful. The prompt should describe what's already in the image, not what you wish were there. "A woman in a red coat standing by a canal at dusk" tells the model what it's looking at so it doesn't "fix" things you liked. Write it wrong (or write a dream prompt instead of a description) and the model will happily invent your dream into the upscale.

The real control:

  • creativity (0.2–0.5, default 0.35) - how much new detail the model is allowed to add while upscaling. The range is deliberately narrow: this route is built to stay close to the source. Nudge it up only if the upscale comes back looking soft and you want a touch more texture.

Plus negative_prompt (steer away from artifacts you keep seeing), seed, and output_format (png/webp/jpeg). Output is a single IMAGE tensor.

Installing it

Part of the "Stability API nodes for ComfyUI" pack - ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Stability-AI/ComfyUI-SAI_API
# restart ComfyUI

No model downloads, requests aside (already on any real install). Key via SAI_API_KEY, a sai_platform_key.txt file beside the node, or the per-node api_key_override - which embeds the key in the workflow, so don't share those files.

Common issues

  • Prompt describes a fantasy, output is fantasy - you're using the prompt wrong. Describe the source image verbatim.
  • Upscale looks smeary on fine detail - this is the trade-off of conservative upscaling on low-res sources; there's no detail to recover at that resolution. Downscale to a sharp base first, or switch to the Creative Upscale node if you actually want detail invented.
  • "payment_required" - credits ran out; it's a metered service.

If you want pixels and only pixels, the honest local answer is still a free ESRGAN pass. This is the same philosophy, hosted, with a slightly generative heartbeat - and for people who can't run anything locally, it's a perfectly good way to get a big, faithful image out of a small one.

CategoryStability

Inputs (7)

NameTypeDefaultDescription
imageIMAGE
promptSTRING
negative_promptoptSTRING
seedoptINT00–4294967294
creativityoptFLOAT0.350.2–0.5
output_formatoptCOMBO3 options: png, webp, jpeg
api_key_overrideoptSTRING

Outputs (1)

NameTypeDescription
IMAGEIMAGE