AuraSRNode

AuraSRNode is a single-input, single-output 4x upscaler from fal-ai's AuraSR, and it is the most hands-off upscale node you'll ever touch. It takes one image, hands you back the same image at four times the pixels, and has exactly zero knobs to turn. No prompt, no CFG, no denoise strength, no model dropdown. That's because it isn't a diffusion model at all - it's a GAN, and it's from 2024, so set expectations before you wire it into your "make everything perfect" pipeline.

What it actually is

AuraSR was released by fal-ai in June 2024 as an open-source "real-world" super-resolution model based on the GigaGAN paper (NVIDIA's scaling-up-GANs work), built on top of lucidrains' unofficial gigagan-pytorch. This node, published by AIFSH the same week the model dropped, was one of the first ComfyUI wrappers for it, and it's never been touched since. The core is a UNet-style upsampler with StyleGAN2-style weight modulation: the input gets a tiny bit of noise added (0.001 scale), runs through a style network and a stack of modulated convolutions and attention blocks, and comes out 4x bigger.

The trick that keeps it cheap is tiling. The model works on 64x64 tiles (input 64, output 256 per the HF config), so whatever resolution you feed it gets padded to a multiple of 64, cut into tiles, processed in batches of eight, and stitched back together. That's why a GAN upscaler with a 2.4GB model runs fine on modest VRAM - it never sees the whole image at once. Where a diffusion-based upscaler like SeedVR2 or SUPIR is restoration that can rebuild a face it doesn't recognize, AuraSR is closer to texture synthesis: fast, plausible detail, no prompt-driven intent.

The one input and the one output

The whole node is:

• image (IMAGE, required) - anything wired from a Load Image or the output of a generation. It upscales every image in the batch, one at a time.
• IMAGE output - the 4x result, ready to feed straight into a Save Image or a VAE decode. There is nothing else to set.

That's the entire interface. If you want a different scale factor, this isn't your node - 4x is the only move.

Install

Two routes, both easy:

ComfyUI Manager - search "AuraSR" and install, then restart.

Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/AIFSH/ComfyUI-AuraSR.git

The README says custom_modules - that's a typo, it's custom_nodes. Then restart ComfyUI. There's no requirements.txt and nothing else to install: it leans on einops, torchvision, huggingface_hub and safetensors, all of which ship with modern ComfyUI. Yes, that's rarer than it should be.

The gotcha that looks like a hang

On the first run the node downloads fal-ai/AuraSR from Hugging Face - about 2.4GB of safetensors plus its config - before the image changes. If your console goes quiet for a while, that's it, not a frozen node. Subsequent runs load from cache in a few seconds.

The model loads onto CUDA with no fallback, so on a CPU-only machine it'll fail; and because the model is a baked GAN output, re-upscaling its output again is a good way to get waxy skin and oversharpened edges. One pass, done.

Where people get burned

The community's verdict on AuraSR, from the launch thread itself: at a big jump like 1024→4096 it produces artifacts and aggressive oversharpening. Scaled down to a more modest effective factor it "looks pretty great." The same logic applies here: it's at its best on small, slightly soft sources getting a moderate boost, not on clean high-res images you want to blow up to print size. The modern advice in the upscaling world - use a fast pixel upscaler if the source is already sharp, a diffusion restorer if it's genuinely damaged - leaves AuraSR in a narrow but real middle lane: it's a fast, zero-config 4x for web-sized content on a low-VRAM card. That lane has been getting narrower since 2025, so grab it while it's still the easy answer, but don't expect it to fix anything.