Replicate topazlabs/image-upscale

Topaz is the company whose desktop apps taught a generation that "upscale" could mean more than bigger pixels - their Gigapixel line is the commercial benchmark people compare everything against. This node is the hosted topazlabs/image-upscale model on Replicate, and its interesting trick is that it exposes the model family as an input, so you're not stuck with one flavor of enhancement. If you're running ComfyUI without a GPU and you want that Topaz-style cleanup on a real photo or a game screenshot, this is the node.

Mechanically it's the usual Replicate schema node: IMAGE in (a tensor, converted to base64), IMAGE out (back as a tensor), plus API_JSON with the payload that went out. The inputs that matter:

• enhance_model - the picker, and where you spend your actual thought. Five choices: Standard V2, Low Resolution V2, CGI, High Fidelity V2, and Text Refine. The rough map: Standard for general photos, Low Resolution V2 for genuinely small/grainy sources, CGI for rendered/game art (it doesn't smear the crisp edges), High Fidelity V2 when you want to preserve detail rather than invent, and Text Refine when the image is mostly text (screenshots, UI) and you want legible glyphs.
• upscale_factor - None, 2x, 4x, or 6x. "None" sounds like a mistake but isn't: it lets the model do its enhancement without enlarging, which is the right call when the source is big enough already and you only want cleanup.
• subject_detection - None, All, Foreground, or Background. Lets the model focus enhancement on the subject instead of the whole frame; useful for portraits and product shots where you want the background left alone.
• face_enhancement (default off), with face_enhancement_creativity and face_enhancement_strength (default 0.8) - Topaz's face pass. Here's the KB warning again: any generative face pass is a rewrite, not a restoration. For a known face, test it at low face_enhancement_strength and check the result actually still looks like the person, because "creativity" is the model's license to invent.
• output_format - jpg (default) or png.
• dry_run / force_rerun - test free, and force a fresh call past ComfyUI's cache.

The thing to internalize, same as every generative upscaler: this is a "more detail" tool, not a "more pixels" tool. The KB's upscaling essay is blunt that for an already-sharp image you want the free ESRGAN/Lanczos rung, not a paid generative model - you'd be paying to have it invent detail that's already there. Reach for this when the source is soft, low-res, or damaged and you want the model to add.

Install is the pack standard - ComfyUI Manager (search ComfyUI-API-DockerCPU), or:

cd ComfyUI/custom_nodes
git clone https://github.com/trustypangolin/ComfyUI-API-DockerCPU
cd ComfyUI-API-DockerCPU
pip install -r requirements.txt

Set REPLICATE_API_TOKEN and restart; it's under 🎨 DockerCPU API/🎨 Replicate.

The gotchas: this is one of the more expensive upscalers in the pack - 4x/6x on big images at high megapixel counts adds up fast, so dry-run, check the API_JSON, and read the invoice after a test batch. And "High Fidelity V2" is not a magic "don't change anything" button - it still rewrites. If you need no change except size, that's the free interpolator's job, not this node's.