Real-ESRGAN Super-Resolution

Real-ESRGAN isn't new and it isn't fancy, and that's exactly the point. If your source image already has the detail you want and you just need more pixels - a photo that's too small, a render you want printable - this is the tool, not a diffusion model. It adds no content, can't hallucinate a face into existence, and runs in milliseconds to seconds instead of the minutes a generative restorer needs. This node is just a clean ComfyUI wrapper around the well-known Real-ESRGAN weights, dropped into Eric Rollei's wider image-processing pack alongside heavier, newer restoration tools.

Why you'd reach for it

The community's read on this hasn't shifted even with SUPIR and SeedVR2 getting all the attention: for pure "more pixels, source already sharp" jobs, the ESRGAN family is still the correct answer and its share of the conversation has actually been rising, not fading, through 2026. It's the free, instant, can't-go-wrong option. You reach for something generative only when the source is genuinely damaged and needs detail invented, not just enlarged.

How it works

Real-ESRGAN is a GAN trained specifically to invert realistic degradation - compression artifacts, sensor noise, mild blur - rather than the clean bicubic downsampling the original ESRGAN was trained on. That's the difference between "ESRGAN" and "Real-ESRGAN": it was built to survive photos that actually came off a phone or the internet, not synthetic test images.

The inputs and outputs that matter

• image - required, obviously.
• model_variant - three choices: RealESRGAN_x4plus for general photos, RealESRGAN_x4plus_anime_6B if your source is anime or illustration, and realesr-general-x4v3 as a lighter, faster option. Pick based on content type, not vibes - the anime variant on a photo (or vice versa) gives you worse output than either would alone.
• scale_factor (1–8, default 4) - the model is natively a 4x upscaler; 1 gets you enhancement without resizing, 8 is stretching it past what the network was trained for and the node's own tooltip calls it experimental.
• tile_size (optional, default 512) - memory management. Bigger tiles are faster and slightly better quality at the seams; smaller tiles let you run on less VRAM. 256 if you're tight, 1024 if you've got room.
• blend_factor (optional, 0–1, default 1) - mixes the enhanced result back with the original. Handy if the model oversharpens or introduces texture you don't want; dial it toward 0 to soften the effect instead of throwing the output away.
• Outputs: enhanced_image to wire into your save/preview node, and processing_info - a string report, useful for debugging what settings actually ran.

How to install it

Search Eric_Image_Processing_Nodes in ComfyUI Manager and install - or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Image_Processing_Nodes
cd Eric_Image_Processing_Nodes
pip install -r requirements.txt

Restart ComfyUI. The pack's actual import-time requirements are light - numpy, opencv-python, scikit-image, scipy, PyWavelets - with torch and friends pulled in as optional extras for the heavier nodes elsewhere in the pack. The Real-ESRGAN weights themselves auto-download on first use (expect the run to pause the first time you hit this node while it fetches them), so budget a bit of first-run patience and disk space.

A word on the pack itself, since it's easy to mistake a young, 10-star repo for abandonware: the author (EricRollei) is an active, known figure in the ComfyUI space - they also maintain nodes for Hunyuan Image 3 and publish models on HuggingFace, and get name-checked in community threads for being fast to iterate. This particular pack is newer and less discussed, but it isn't a one-off.

Common issues & troubleshooting

Out of VRAM on large images. This is the standard tiling trade-off, not a bug - drop tile_size to 256 before you drop resolution or give up.

Output looks over-sharpened or plasticky. Common GAN-upscaler artifact, especially on already-clean sources. Pull blend_factor down toward 0.5–0.7 rather than fighting it with post-processing.

Anime output looks smeared, or photo output looks cartoonish. You've got the wrong model_variant for your content - this is the single most common misuse of any Real-ESRGAN wrapper, in or out of ComfyUI.

First run hangs or is slow. That's the model weights downloading, not a freeze. Check your network and give it a minute before assuming something's broken - this is generic to every node in the pack that pulls a pretrained checkpoint on first use, not specific to this one.

Missing custom node errors on a shared workflow. If you loaded someone else's graph and ComfyUI can't find this node, install the pack first (above), then reload - standard ComfyUI custom-node friction, not this pack being unusual.