Replicate jingyunliang/swinir

If you've ever googled "SwinIR ComfyUI," this is the node you landed on. It's the classic Swin Transformer image-restoration model - from the paper Activating More Pixels in Image Super-Resolution Transformer - packaged as a Replicate model and dropped into ComfyUI by the Comflowy pack. SwinIR has been a community staple since 2021 as the go-to "clean up a real photo" upscaler, so if you're here, you probably have a blurry or noisy image and want it fixed without a doctorate in model config.

The honest framing first: SwinIR is open source. You can run it locally for free if you have a GPU. What this node buys you is the convenience of not hunting down the right weights, and the ability to call it from any machine - including one with no GPU at all. Comflowy runs it on Replicate's servers and bills you in credits. For the occasional restoration it's fine; for heavy batch work, the local version wins on price.

How it works

Every Replicate node in this pack is auto-generated from the model's OpenAPI schema and proxied through Comflowy's backend. Your input image gets base64-encoded, POSTed to app.comflowy.com/api/open/v0/flowy, and Comflowy's service runs the prediction on Replicate with your Comflowy API key doing the auth. The result URL comes back, the node downloads it, and converts it to a normal ComfyUI IMAGE tensor you can wire anywhere - save, preview, feed into another upscaler.

One subtle thing: results are cached by default. The node only re-runs when its inputs change. If you tweak nothing but want a fresh pass - say, the server hiccuped - flip the force_rerun boolean and it'll fire again.

The inputs that matter

• image (required) - your input, any standard image.
• task_type - the job SwinIR should do. Default is Real-World Image Super-Resolution-Large, which is what most people want. Also available: Real-World Image Super-Resolution-Medium, Grayscale Image Denoising, Color Image Denoising, and JPEG Compression Artifact Reduction.
• noise - only active for the two denoising tasks. Pick 15, 25, or 50 matching how noisy the source is.
• jpeg - only active for JPEG artifact reduction; 40 is the default and sensible for most compressed images.

Output is a single IMAGE. That's it - this is a one-in, one-out node.

Installing it

You're not installing this node alone; you're installing the whole Comflowy pack, since the Replicate nodes ship inside it. The easy way:

1. Install Comflowy's Custom Nodes via ComfyUI Manager (search "Comflowy" in Manager's install tab).
2. Or, in your terminal: cd ComfyUI/custom_nodes && git clone https://github.com/6174/comflowy-nodes
3. Restart ComfyUI. The only real dependency is requests, which you almost certainly already have.

Then add a Comflowy Set API Key node, paste the key from your comflowy.com account settings (avatar → Settings → API Key), run it once, and you can delete it. No key, no Replicate node works - you'll get a "API Key is not set" error.

Common issues

• Network errors. Comflowy's nodes are just HTTP calls. If you see Failed to get response from LLM model with https://app.comflowy.com/api/open/v0/prompt or a plain request failure, that's the README's own warning: check your network and your route to the API host - restricted regions and flaky proxies are the usual culprits.
• Credits. Every run costs credits, and unlike most of the pack this one needs no API of its own - it's all billed through Comflowy. Watch the counter if you're running batches.
• Wrong results per task. If you set a denoising task, the noise field matters; if you set JPEG reduction, jpeg matters. Pick the task that matches your actual problem and the others become irrelevant.

If you just want more pixels out of an already-clean image, this is arguably overkill - an ESRGAN model locally is free and instant. But for restoring a genuinely damaged photo from a laptop with no GPU, this node is the low-friction path.