Clarity AI Upscaler

If you dropped this node into a workflow expecting a free local upscaler, stop and read this first: it doesn't run anything on your GPU. Clarity AI Upscaler is a thin client for clarityai.co's cloud service - your image goes out over the wire, their servers do the actual work, and you get a bigger, more detailed image back. On a platform like ComfyICU that's an interesting quirk: the GPU tier you picked for the run barely matters for this node, because the heavy lifting happens on Clarity's infrastructure, not the executor.

It sits in the "creative upscaler" family alongside Magnific AI and, on the open-source side, SUPIR and SeedVR2 - tools that don't just add pixels, they invent detail that was never in the source: skin texture, fabric weave, foliage. philz1337x, the node's author, also maintains a separate repo called clarity-upscaler that reverse-engineers a chunk of that same idea locally and for free, using tiled MultiDiffusion plus ControlNet and LoRA over an SD1.5 checkpoint. That repo made real noise on Reddit back in 2024 as "the free answer to Magnific," and one commenter's blunt summary still holds up: it's about the only affordable way to get this kind of output resolution with a generally good result, cheaper than Magnific, with the catch that it can get grainy and drift toward "Automatic1111 land" at higher multipliers. This node's own README even points at a free-workflow.json that leans on that same lineage and gets close to the paid results - worth trying first if you don't want a subscription.

The node itself is simple. Feed it an image, optionally a prompt, and four sliders that mirror the classic Magnific-style dials: creativity, resemblance, dynamic, and fractality, each ranging -10 to 10 and defaulting to 0. Push creativity up and the model takes more liberty reinterpreting the image - more hallucinated detail, less faithful to what was actually there. Push resemblance up to pull it back toward the source structure. dynamic and fractality behave like the contrast/texture-density knobs the rest of this category uses; the README doesn't document exact behavior beyond the names, so treat them as "nudge and compare" rather than something to reason about precisely. style switches between default, portrait, and anime presets, and scale_factor sets your multiplier from 2x up to 16x. There's one output, IMAGE, which goes straight into a Save or Preview node - or into a separate face-detail pass if you pushed creativity hard on a portrait.

Installing it is the usual two-path story: open ComfyUI Manager, search "Clarity AI," install, restart. Or manually, cd custom_nodes && git clone https://github.com/philz1337x/ComfyUI-ClarityAI and restart. Either way, you're not actually done - this node needs a Clarity AI API key before it'll run. Grab one at clarityai.co/comfyui, then either set it as the CAI_API_KEY environment variable or drop it into a cai_platform_key.txt file inside the node's own folder. There's also an api_key_override field on the node for a quick test, but the README is explicit that you should clear it before sharing a workflow - otherwise you're handing your key to anyone who opens the JSON.

The most common "why isn't this working" isn't a bug, it's the business model: Clarity AI is not free, and the node gives you no hint of that until a call fails. It runs on a token system, with reported tiers from a one-time ~$7 bucket up to roughly $29–$74/month for heavier use, so if the node errors out or returns nothing, check your token balance before you go hunting through logs. Second, because every run is an outbound API call, a bad or expired key and a genuine network hiccup look identical from inside ComfyUI - both just show up as a failed or hanging node - so it's worth confirming your key and clarityai.co's own status before assuming your workflow is broken. And third, the general caution that applies to this whole category: crank creativity high on a face you actually recognize and you'll get a face that isn't quite that person anymore. These models don't preserve identity, they invent a plausible one from the prompt and the source's overall shape. For a faithful restoration, keep creativity low and resemblance high, or give the face its own separate pass.