🐳RealESRGAN 超分辨率

RealESRGAN is the workhorse of the "make it bigger" layer - a classic super-resolution model that's been the community's default for years, especially the anime-tuned variants. This node wraps a 4x RealESRGAN in ONNX Runtime and gives you exactly two knobs: the image and a blend dial. From 🐳 Pond Nodes (comfy_Pond_Nodes), it's the pack's "I need this bigger and I don't want to think about it" node.

The name is a lie in one important way: it's not the PyTorch realesrgan package doing the work. It's an ONNX model running through onnxruntime, which is faster and lighter than the full Real-ESRGAN stack - but it's also what creates the trap below.

How it works

The image goes into an ONNX inference session, GPU first (CUDAExecutionProvider) with CPU fallback, and comes out 4x larger. Then the clever part - the blend slider. At 1.0 you get pure RealESRGAN output. Below that, the node bicubically upscales the original to match and blends the two: at 0.5 you're seeing half RealESRGAN, half plain interpolation. That's the standard trick for taming RealESRGAN's tendency to oversharpen and plasticize skin - a little blend trades absolute sharpness for a much more natural look.

The inputs and outputs

• image - what you're upscaling.
• blend (0–1, default 1) - how much RealESRGAN vs. plain upscale. The single dial worth touching; 0.4–0.7 is a common range for photos.

One output: IMAGE, 4x the input dimensions. Wire it straight into your upscale pass - before a hi-res fix, or as the final output stage.

Installing it - and the model path trap

Standard pack install (Manager search comfy_Pond_Nodes, or clone + pip install -r requirements.txt), and this one genuinely needs onnxruntime from that file. But here's the trap: the README says to drop the model in ComfyUI/models/upscale_models/. The source code doesn't look there. It hardcodes the path to ComfyUI/models/real_esrgan_x4_fp16.onnx - the root of the models folder, a specific filename. Follow the README and you'll get a "找不到模型文件" (model not found) error every time.

So, the working setup:

# put the file at this exact path:
ComfyUI/models/real_esrgan_x4_fp16.onnx

Grab a real_esrgan_x4_fp16.onnx from a RealESRGAN ONNX release (the x4plus family), name it exactly that, and restart. No upscale_models folder involved.

The gotcha

Beyond the path: ONNX RealESRGAN is a fixed 4x model - there's no scale selector, no tile size, no model variant dropdown. Big images can eat surprising amounts of VRAM, and if you hit OOM, the ONNX session doesn't automatically tile. Also, blend only softens the upscale; it won't fix a mushy source. For a sharp-but-natural look on portraits, the recipe is: blend ~0.5, then a light sharpen afterward. If you want the full control panel (tiling, model choice, the anime-6B variant), this isn't it - that's a job for a purpose-built upscale node suite.