Where Do chaiNNer AI Upscaling Models Come From?

One of the most important things to understand about chaiNNer is that the application itself is not the upscaling model. chaiNNer provides the node system, runtime and processing pipeline. The neural network you load into that pipeline determines how the image is reconstructed.

chaiNNer is the engine. The model decides the result.

A useful mental model is:

The complete path looks like this:

Swap the model and the same chain can produce a visibly different result. One model may preserve realistic textures well, another may be better for illustration, and another may aggressively remove compression artifacts.

What is a .pth model file?

Many super-resolution models used with chaiNNer are distributed as files ending in .pth. These are PyTorch model weight files. They contain learned parameters from a neural network that has already been trained.

The file is not an application and it does not run by itself. chaiNNer loads it through a compatible model loader and passes the resulting model object to an upscale node.

During normal use, you are not training RealESRGAN. You are running inference with a model that has already been trained elsewhere.

Our starting model: RealESRGAN_x4plus

For our first chaiNNer workflow, we used RealESRGAN_x4plus.pth. It is a 4× ESRGAN-family model designed for practical general image restoration and upscaling.

Its native scale is 4×, which means an image such as 1000 × 1000 pixels becomes approximately 4000 × 4000 pixels when the model is used at its native scale.

The official Real-ESRGAN project provides pretrained models including RealESRGAN_x4plus, RealESRGAN_x2plus and anime-oriented variants.

Where do these models come from?

There are three common sources:

1. Official model repositories maintained by the original project authors.
2. Model databases that catalog architectures, intended uses, licenses and download sources.
3. Community-trained models created for specific types of artwork, restoration or compression problems.

For well-known models, the best starting point is normally the original project repository and its official release files. For Real-ESRGAN, that means the official xinntao/Real-ESRGAN GitHub repository →

OpenModelDB: a practical model catalog

OpenModelDB → is one of the most useful places to research image-restoration and super-resolution models.

A model page can expose information such as:

• architecture,
• native upscale factor,
• model size,
• intended use,
• author or source,
• license or usage rights,
• download location.

For example, the OpenModelDB entry for RealESRGAN_x4plus identifies it as a 4× ESRGAN model for general restoration and links the PyTorch .pth file to its GitHub-hosted source.

Different models have different specialties

There is no single universally best upscaling model. A strong model choice depends on the source material.

The official Real-ESRGAN family itself illustrates this. RealESRGAN_x4plus is a general-purpose baseline, while RealESRGAN_x4plus_anime_6B is a smaller model optimized specifically for anime-style images.

This is why building a small tested model library is more useful than collecting hundreds of files with no idea what they are for.

Licensing matters — especially for commercial workflows

When a model is used in a commercial production pipeline, licensing must be checked separately from technical compatibility.

The model file, the codebase and the artwork you process can all have different rights attached to them. Do not assume that every community model you find online has the same permissions.

The OpenModelDB entry for RealESRGAN_x4plus lists BSD-3-Clause rights and indicates commercial use, redistribution and modification rights under that license. Other models may use different licenses or have additional requirements.

Before adding a model to a commercial pipeline, check:

• the original author or repository,
• the stated license,
• whether commercial use is allowed,
• whether attribution is required,
• whether redistribution of the model file is allowed.

A safer workflow for downloading models

This avoids a common mistake: building a folder with dozens of anonymous model files and later having no idea which one is safe, licensed or appropriate for a given asset.

Why we started with RealESRGAN_x4plus

We needed a practical baseline: a model that is well documented, easy to identify, compatible with common chaiNNer super-resolution workflows and useful for general artwork.

RealESRGAN_x4plus fits that role well. It is not presented here as the best model for every image. It is useful because it gives us a stable reference point.

Once a baseline exists, testing becomes more meaningful:

Build a tested model library, not a download archive

A useful model library should record more than file names. Track what each model is good at, what it damages, native scale, architecture, license and the kinds of images used during testing.

A simple table is enough:

That turns model selection from guesswork into a repeatable production decision.

Final takeaway

chaiNNer gives you the pipeline, but the model determines much of the visual result. Understanding where models come from, what they were trained to do and what license they use is part of using AI upscaling responsibly.

Start with a well-documented baseline, test on your own reference images, keep notes and only expand your model library when a new model solves a specific problem better than the one you already have.