July 22, 2026·7 min read·app.nz

ComfyUI lesson 6: Use LoRAs without losing the base model

Balance model and CLIP strength, compare with a fixed seed, and avoid adapter pileups.

Listen to this article

On-device voice

Uses the voice built into your browser; no article text leaves this page.

Audio narration is not supported by this browser.

A LoRA is a small adapter that nudges selected layers of a base model. It can add a style, character, object, or behavior without replacing the checkpoint. Small file does not mean “safe at any strength.”

Wire the adapter correctly

Put Load LoRA between the checkpoint and both downstream branches. Feed its MODEL to KSampler and its CLIP to the prompt encoders. If only the model path is adapted, trigger words and style semantics may behave differently from the LoRA author's examples.

Start with model strength 0.6 and CLIP strength 0.4–0.6. Fix the seed and render:

no LoRA (strength zero);
0.4;
0.7;
1.0.

Look for the first point where the desired feature appears, then stop. Overshooting often causes crunchy edges, repeated motifs, distorted anatomy, or the same face in every composition.

Compatibility is not optional

An SD 1.5 LoRA does not belong on SDXL. A FLUX adapter does not belong on Proteus. Match architecture, read the source license, preserve the original filename, and record the source URL. The app.nz model linker accepts public Hugging Face files and Civitai versions and can warm a shared content-addressed cache before the worker starts.

The runnable lesson uses Proteus V0.4 plus one detail adapter and exposes LoRA strength as an API input. After the single-adapter experiment, try two LoRAs only if each one works alone. Blend at modest strengths; adapter stacking is a composition problem, not a contest to reach 1.0.

Run the lesson, do not just read it

Open the tested workflow in Comfy Studio
Download the exact API-format JSON
CLI: app comfy download proteus-v04-lora

The gallery records model sources, minimum VRAM, custom-node requirements, and verification state beside the graph. That is the portable unit for this series.

Series progress: lesson 6 of 10. Previous: Build image-to-image and master denoise Next: Control composition with ControlNet

Wire the adapter correctly

Start with model strength 0.6 and CLIP strength 0.4–0.6. Fix the seed and render:

no LoRA (strength zero);

0.4;

0.7;

1.0.

Look for the first point where the desired feature appears, then stop. Overshooting often causes crunchy edges, repeated motifs, distorted anatomy, or the same face in every composition.

Compatibility is not optional

Run the lesson, do not just read it

The gallery records model sources, minimum VRAM, custom-node requirements, and verification state beside the graph. That is the portable unit for this series.