Load AuraSR

Load AuraSR does exactly one job, and it's the job nobody should have to think about: it takes a downloaded AuraSR weight file and hands you a ready-to-run model object. That's it. One dropdown, one output, no sliders to misconfigure. If you've ever used Load Checkpoint, you already know this node - it's the same shape, just for a GAN upscaler instead of a diffusion backbone.

The context helps. AuraSR is fal's open 4x upscaler, a reproduction of the GigaGAN paper that shipped in mid-2024 and immediately split the community on quality (more on that in the Run AuraSR article). This pack wraps it in two nodes, and this is the loading half. You'll never wire Load AuraSR's output to a viewer or a save node - the AURASR_MODEL output is a Python object, not an image, and it feeds exactly one thing: the AURASR_MODEL input of Run AuraSR. Think of this node as the cable you plug in once, then forget.

How it works

Under the hood the pack embeds fal's aura_sr.py and its config.json directly, so there's no separate fal Python package to install. When you execute the node it grabs the file path from ComfyUI's upscale_models folder, loads the safetensors into memory on CUDA, and returns the model.

The one input that matters is model, which is an enum of every .safetensors file ComfyUI finds in models/upscale_models, defaulting to aurasr.safetensors. Two notes here. First, the filter is strict - a .ckpt dropped into the same folder will not appear in the dropdown, even though the underlying loader could read it. Second, because it lists everything in that folder, your entire ESRGAN collection shows up too; pick the AuraSR file, not 4x-UltraSharp.

Installing it

Install the pack normally, then grab the weights - that's the part people forget:

cd ComfyUI/custom_nodes
git clone https://github.com/alexisrolland/ComfyUI-AuraSR.git
cd ComfyUI/custom_nodes/ComfyUI-AuraSR
# on Windows with the bundled python:
..\..\python_embeded\python.exe -s -m pip install -r requirements.txt

The requirements.txt (torch, torchvision, einops, Pillow, requests) is mostly a formality - ComfyUI already ships all of it, and the one thing that looks missing, safetensors, is bundled too. The real requirement is the model download. Per the README, fetch from the Lexius/AuraSR repo and drop the files into ComfyUI/models/upscale_models:

• aurasr.safetensors (v1, ~2.5 GB)
• aurasr_v2.safetensors (v2, ~2.5 GB)

Or skip the README's typos and just use ComfyUI Manager, search "AuraSR", and let it clone the pack for you. The model download is still manual either way.

Common issues

• The dropdown is empty, or "Model file not found." The files are in the wrong place. They belong in models/upscale_models, not models/checkpoints - the node only reads the upscale_models folder.
• Nothing shows but my ESRGANs. That's the folder filter doing its job; your AuraSR file either isn't downloaded yet or isn't named with a .safetensors suffix. The README's link text says aurasr.safetensor (missing the s) - the real filenames are aurasr.safetensors and aurasr_v2.safetensors.
• Slow first run. You're loading a 2.5 GB model into VRAM; the first execution of any workflow will stall for a few seconds. That's loading, not a hang.
• It's GPU-only. The pack hardcodes CUDA - no CPU fallback. If you're running ComfyUI on a CPU box, this node will error, not fall back.

If you have both v1 and v2 downloaded, pick v2. It's the same interface, and it fixes a lot of the v1 quality complaints.

Want the model that runs? Wire Load AuraSR straight into Run AuraSR and you're done - that's the whole pack.