Run AuraSR

Let's be honest about AuraSR up front, because it'll save you an hour of disappointment: it's a GAN upscaler, not a diffusion upscaler, and that's the whole story. It's the "more detail" kind of upscaler - it invents plausible texture that wasn't in the source - but it does it in a single fast forward pass instead of many denoising steps. When fal released it in mid-2024, the community verdict was split down the middle: "works amazingly fast" from one person, "this is garbage" from the next, usually about the same image. The truth is that AuraSR is a very specific tool. It shines on clean, uncompressed sources at moderate sizes, it's quick, and it's light on VRAM. It gets ugly the moment you feed it JPEG artifacts or try to jump a huge image in one hop.

That's where this node sits: it's how you run AuraSR locally in ComfyUI, the opinionated companion to the pack's Load AuraSR. You use it when you want 4x with plausible detail and you want it now, or when your GPU can't fit a diffusion upscaler like SUPIR or SeedVR2. If you just want the image bigger and it's already sharp, honestly reach for a plain 4x ESRGAN - it can't hallucinate anything. This node is for when you want it to invent detail.

How it works

Run AuraSR takes the model object from Load AuraSR plus any image tensor, then does what the model does: it tiles the input at the model's native resolution (64px input tiles per the pack's config), runs each tile through the GAN, and stitches the results back together. That internal tiling is why it handles images of any size without exploding your VRAM.

The inputs that matter

There are only three, and only one is interesting:

• AURASR_MODEL - the output of Load AuraSR. Nothing else goes here.
• IMAGE - any image tensor. From a Load Image node, from the KSampler, from whatever. You can upscale a generation directly in the same graph.
• avoid_seams - default true. This is the interesting one. On, the node runs the image through twice with offset tiles and blends the passes using checkerboard weights to kill visible tile seams. Off, it's a single tiled pass - faster, but large images can show tile boundaries. Leave it on unless you're upscaling small images where seams aren't a problem and you want the speed.

The output is a single IMAGE tensor - wire it to Save Image or a preview node and you're done.

What actually goes wrong

The failure modes are the model's, not the node's. Community testing at launch was loud about all of them:

• Compressed sources look bad. AuraSR has no error correction. Feed it a JPEG and you get haloing, oversharpening, and a dithering noise on flat surfaces. Clean PNGs are its happy place. One workaround people landed on: pre-upscale the compressed image with a traditional ESRGAN model, then run AuraSR.
• Big single-hop upscales fall apart. A 1024×1024 render upscaled straight to 4096×4096 came back with artifacts and a weird "texture" people memorably described as having been "trained in dollar bills." If your source is large, downscale first or use a different tool.
• It's 4x only. No scale factor slider, no 2x option. What you see is what you get.
• It's quiet now. By 2026, AuraSR has mostly faded from the conversation - SeedVR2 took over the "add detail" slot. This node still wins on speed and VRAM if that's your constraint, and it's worth keeping around. But if you found this article while googling "best upscaler," read the KB's upscaling essay first; AuraSR is a tool, not the answer.

If you have the choice, run aurasr_v2.safetensors - v2 fixed a lot of v1's compression-artifact issues, and it's the same node, same wiring, better results. Load AuraSR → Run AuraSR → Save Image is the entire workflow, and it takes about five minutes to set up.