HitPaw General Image Enhance

There are two very different jobs hidden under the word "upscale," and this node does the generative one. HitPaw General Image Enhance takes a low-res image, sends it to HitPaw's hosted API, and gets back a version that's bigger and has artifacts and noise scrubbed out. Think of it as the cloud version of a SUPIR-style restorer - same family as the local upscalers the community argued about for years, but running on someone else's GPU and billed per call.

It ships with ComfyUI core as a partner node, so there's nothing to install. You trade the "free local" part of upscaling for the "it just works on any machine" part: no model download, no VRAM pressure, no tiled-diffusion dance to fit your card. If you're upscaling a handful of hero images and don't want to babysit a 16GB VRAM tiling workflow, this is a reasonable shortcut.

How it works

The node uploads your image, creates a task on HitPaw's photo-enhancer endpoint through Comfy's proxy, then polls until the job finishes and downloads the result back as a normal IMAGE tensor. Two model flavors exist: generative for general images and generative_portrait for faces - the portrait model is tuned to actually keep people looking like people while it cleans them up.

The inputs that matter

• image - what you're fixing. Wire in anything that outputs IMAGE: a generated render, an old scan, a downscaled still.
• model - generative or generative_portrait. If the subject is a face, the portrait model is usually the better pick; it's also the cheaper one.
• upscale_factor - 1x, 2x, or 4x. Yes, 1x is a legit option: it runs the denoise/artifact pass without enlarging, which is often exactly what a crusty source needs.
• auto_downscale - defaults off, and this is the one that bites. The service caps output at 32 megapixels. Feed it a big input and request 4x and you'll get a hard error unless you flip this on, in which case the node quietly shrinks the input first so the output fits.

The single output is IMAGE, which you can save, compare against the source with a Mask node, or push into further processing.

Where people get burned

• The 32MP cap. It's right in the node's own description, but everyone discovers it by hitting the error first. If you're going big, turn on auto_downscale and let the node do the math.
• Batch of one. This isn't a batch upscaler; each call handles a single image, and each call is a separate charge. For a hundred-frame batch you'll want a local upscaler, not this.
• The poll. Cloud jobs take time - the node literally polls the task status until it's done. A 4x upscale of a large image isn't instant, so don't panic when the node sits there "running."
• Cost. Per-call pricing, shown on the price badge. The portrait model costs less than the general one, which is a nice coincidence given it's usually the better fit for faces.