Skip to content
CozyGPU

Custom LoRAs

Drop a .safetensors file on the app. It shows up in the LoRA picker next to the built-in ones.

Adding a LoRA

  1. Drag a .safetensors file onto the app window, or use Add LoRA in the picker.
  2. Pick the model family it was trained for, give it a display name, and add its trigger words if it has any.
  3. Set a default strength (0.8 is a sensible start; the slider goes to 1.0).

The file is hashed and stored locally as loras/<sha256>.safetensors in the app's data folder, so the same file is never stored twice. It is then copied to each of your RunPod network volumes in the background; the picker shows a sync badge until every volume has it.

Which models

Custom LoRAs work on all three models: Z-Image Turbo, Qwen Image 2.1 (text to image and image edit) and MiniMax H3 (text to video and image to video). Each LoRA belongs to the model family you pick when you add it, and is chained into that model's workflow the same way the built-in LoRAs are. A job takes up to three LoRAs on the image models and up to two on MiniMax H3.

Requirements

  • The RunPod S3 API key must be set; the app needs it to copy the file to your volumes.
  • The first LoRA you add creates the network volumes (10 GB each, in up to three datacenters near the GPU pools the app uses). RunPod bills those per GB per month.
  • Only LoRAs trained for the model you select will work well. A LoRA trained for a different base model loads but produces garbage or an error.

Using it

Select the LoRA in the picker, adjust strength, and put its trigger words in the prompt. Several LoRAs can be stacked; order and strength both matter. The built-in realism and film LoRAs make good partners for a style LoRA.

Removing

Delete from the picker. The app removes the copy from every volume, the database row, and the local file once no other entry refers to it.