replicate/qwen-image-2.1-lora

Run Qwen Image 2.1 with your own custom LoRAs

Public
3 runs
Weights
Playground API Examples README Versions

Run time and cost

This model runs on Nvidia H100 GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

Qwen Image 2.1 with LoRAs

Generate images from a prompt, or supply reference images and describe the edit you want. Use the base model on its own, or bring a Qwen Image 2.1 LoRA for your own character, subject, or style.

This endpoint is for inference. To train a LoRA, use SimpleTuner’s Qwen Image trainer.

Getting started

  • Enter a prompt. For editing, add an image and, optionally, reference_images.
  • To use a standard SimpleTuner-trained Qwen Image 2.1 LoRA, put a publicly accessible direct .safetensors download URL in lora. Include its trigger word in your prompt if it has one.
  • Start with lora_scale at 1. Leave lora empty to use the base model, or add a second adapter with lora_2.

The defaults are 40 steps and guidance 1. Choose an aspect ratio and get PNG or WebP images back. LoRAs must match Qwen Image 2.1, not the original Qwen Image.

Billing and faster starts

This runs on one H100, billed by the second. See Replicate pricing.

For regular use, create a deployment and set its minimum instances to 1. Keeping a worker warm avoids repeated cold starts and makes response times more predictable. It still costs money while idle; a minimum of 0 allows scaling to zero, but cold starts can return.

Model created