Exactly Illustrative Training

Training a LoRA is usually a local hardware commitment: hours of VRAM, learning-rate babysitting, and a setup that makes beginners bounce off the hobby entirely. Runware's Exactly Illustrative Training is the opposite proposition - you hand it a model to train and a ZIP of images, it runs the whole training job in the cloud, and you get the trained result back as a path. No local GPU involved, no watching your fans spin.

It's a training node with a narrow shape. You're not picking 500 hyperparameters - the widgets are a model reference and a dataset, and the rest is Runware's job. That makes it ideal for the classic beginner use case: "I want the model to know this character / this style," with a handful of clean images rather than a full data pipeline. If you've read the community's LoRA-training lore and want none of the pain, this is the shortcut.

What you set

• importModel (required) - wired from the Runware Import Model builder node. That builder holds the real model metadata: air (the provider:model@version identifier of the base model you're training on), name, whether it's private, plus optional description/version fields.
• dataset (required) - a ZIP of training images as a UUID or URL. JPEG, PNG, and WebP are supported; images are automatically converted to JPEG and downscaled to 4096px on the long side, and each image caps at 50MB. So: curate the images, zip them, host the ZIP somewhere reachable, paste the URL.

Output is file_path, a STRING - the path to the trained model once the job finishes, saved into your ComfyUI output folder. You'd feed that to a loader or a file-based consumer downstream.

How it works

This node uses taskType: training instead of the inference tasks the rest of the pack runs. Same plumbing - request over REST through the Runware SDK - but the job is a training run, not a generation. It also doesn't return an IMAGE or AUDIO tensor; training results land on disk and the node hands you the path. Because it's billed by the job, the title bar's cost readout matters more here than anywhere else in the pack - check it before you fire a big dataset.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

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

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login. This is a paid cloud API - training jobs cost real money, not just time.

Troubleshooting

Most failures here are dataset problems before they're model problems. A ZIP that won't download, a file over 50MB, or images in an unsupported format will bounce the job - check those first. If the trained result comes out underfitted, the fix is usually more, cleaner images with consistent framing, exactly as the local-training community will tell you: garbage in, LoRA out. And if the import model isn't recognized, re-check the air field in the Import Model builder - that identifier is case-sensitive and versioned.