Patreon
Home / Guides / Qwen Image LoRA Training Locally, Step by Step With Musubi T

Qwen Image LoRA Training Locally, Step by Step With Musubi Tuner

By Furkan Gözükara · Updated Nov 21, 2026

You can do Qwen Image LoRA training locally on a Windows PC with as little as 6 GB of VRAM using the SECourses Musubi Tuner app, a 1-click interface for Kohya Musubi Tuner with ready presets for every GPU size. You pick the Qwen model, let the app load the preset for your card, point it at a folder of images captioned with a single trigger word, and train; the same app also runs on RunPod, SimplePod and Massed Compute.

What you need

ItemDetails
GPUNVIDIA, from 6 GB VRAM for both LoRA and full fine-tuning
AppSECourses Musubi Tuner, available to members on Patreon
PlatformsWindows, RunPod, SimplePod and Massed Compute
PythonThe installer creates its own Python 3.12 virtual environment
LibrariesPyTorch 2.13 with CUDA 13 and precompiled FlashAttention, SageAttention and xFormers
Dataset28 images worked in the tests; more angles work better

Before the first install, set up Python, Git, FFmpeg, CUDA and the C++ tools once with the Windows requirements tutorial, or use the 1-click AI requirements installer.

Which Qwen model to train

The Qwen Image card in the app has a Model Version setting that switches between the text-to-image model and the edit models. The Qwen Image 2512 training comparison trained four of them and compared the results:

  • Qwen Image 2512: the best model to train right now. Against the older base model it gave a massive quality boost with less noise. In the first Qwen Image 2512 training test it behaved like FLUX Dev: more training made it more realistic, with less noise.
  • Qwen Image (older base): showed overfitting and quality degradation at 240 epochs.
  • Qwen Image Edit 2509 and 2511: for teaching edits and styles. Edit models need fewer epochs, and overtraining changes how faithfully they keep the input image. For style transfer, 1328x1328 was the sweet spot.

The app also lists Qwen-Image 2.1 and 2.1 Turbo among its 33 trainable base models.

Step by step

  1. Install. Download the zip from the SECourses Musubi Tuner post, extract it into a short folder path without spaces and double-click Windows_Install_and_Update.bat. It installs with uv; a fresh install took 144 seconds with the package cache already filled. Run the same file later to update.
  2. Download the models. Run Windows_Download_Training_Model_Files.bat and type the number of the Qwen bundle. Files download over 16 parallel connections, are checked by SHA-256 and resume if interrupted. No Hugging Face login is needed.
  3. Start the app with Windows_Start_App.bat; it opens in your browser.
  4. Pick the model and preset. Choose Qwen Image and keep LoRA. DoRA is the default and loads directly in ComfyUI and SwarmUI. The app detects your GPU and loads its measured preset.
  5. Set the dataset. Put your images in a folder named with a repeat count and your trigger word, place it inside a parent folder, and select the parent. Click Check dataset to see the steps per epoch and any problems before training.
  6. Turn on the Best Checkpoint Finder, which scores each saved checkpoint for likeness, prompt following and over-training and tells you which one to use.
  7. Start Training and follow the Live Training Monitor: loss, speed, ETA and GPU memory.

The 0 to hero Qwen Image training tutorial shows the whole process in about 1 hour 52 minutes, including how to resume an incomplete LoRA training.

Dataset and captions

The rules below come from the Stage 1 research, more than 35 Qwen LoRA trainings, in the Qwen Image LoRA Stage 1 results post, and from the tutorial:

  • Caption only ohwx. Nothing else, not even a class token. The tutorial has a chapter on why a single trigger word works best.
  • A small dataset can work. A weak 28-image dataset worked for a person, which the post calls the hardest case. A dataset with more angles performs better.
  • Add variety. Vary clothing, backgrounds and angles. The tutorial auto-crops and resizes images with the Ultimate Batch Image Processing tool before training.
  • Style and product datasets differ. For a style, keep it consistent with no repeating elements; for a product, include correct proportions and detail shots.

If you want detailed captions anyway, the app's Image Captioning tool writes them with Qwen2.5-VL 7B on your own GPU.

How many epochs

The first Qwen Image 2512 test used the 200 epoch Tier1_29400_MB.toml config on 28 images:

EpochsSteps (28 images)Result
1203,360Noisiest of the three
1805,040Less noise
2406,720Most realistic, least noise

The full-quality comparison is linked from that post. For the edit models, plan fewer epochs; the comparison video tested them at 120 and 180 epochs. The Musubi presets also include 66 Qwen research presets in 50, 100 and 200 epoch versions, and saving a checkpoint every few epochs lets you compare several points from one run.

Presets, VRAM tiers and speed

Preset file names carry the GPU tier and the measured VRAM peak, and every preset keeps at least 1.5 GB of its GPU free. The tutorial compares Tier 1 (BF16) with Tier 2 (FP8 Scaled) output quality, and shows how to convert a 40 GB fine-tuned Qwen model to FP8 Scaled after training. In the app's own benchmark, Qwen-Image 2512 trained 24.5% faster than stock Musubi Tuner at the same quality, and Qwen-Image-Edit 2511 40% faster. For more on fitting training into small cards, see training a LoRA on a low VRAM GPU.

Training edit models

For Qwen Image Edit 2509 and 2511, put the reference images next to each target image in a sub-folder; Check dataset pairs every target with its references so you see exactly what the model learns. The tutorial also shows how to teach the edit model new commands with control images and what that does to performance.

Generating with your Qwen LoRA

Load the saved .safetensors file in SwarmUI or ComfyUI like any LoRA. The Stage 1 images were made in SwarmUI with the Qwen Image 8 Steps Ultra Fast preset, which uses the 8-step Lightning LoRA, plus the trained LoRA at strength 1. Upscale Images 2X then took them from 1328x1328 to 2656x2656. For even more realism, the Qwen realism presets tutorial adds High Realism and Ultra Realism presets. Setup is covered in the SwarmUI install guide.

Why train Qwen instead of FLUX

In a Qwen Image vs FLUX SRPO comparison on the same 28-image dataset, Qwen was far ahead at complex prompts and emotions. One technical difference: Qwen Image uses an exponential timestep shift, which at its native 1664x928 resolution equals a linear shift of 2.205, as explained in the timestep shift article; you do not need to understand it to train. If you still want FLUX, see the Kohya FLUX LoRA settings guide.

FAQ

Can I train a Qwen Image LoRA on a 6 GB GPU?

Yes. The tutorial and the Stage 1 configs train both LoRA and full fine-tuning on Windows GPUs with as little as 6 GB of VRAM. The app detects your GPU and loads the preset measured for it.

Should I add a class word like "man" to the captions?

No. The Stage 1 recommendation is to caption with ohwx only, without a class token, and the tutorial explains why a single trigger word works best. Research into a better activation token than ohwx was planned for the next stage.

Which is better to train, Qwen Image 2512 or Qwen Image Edit 2511?

They do different jobs. Qwen Image 2512 is the text-to-image model and gave the best training results in the comparison. Qwen Image Edit 2511 learns edits and styles from target and reference images, and needs fewer epochs.

Can I train Qwen Image in the cloud instead?

Yes. The zip includes installers and instructions for RunPod, SimplePod and Massed Compute; on Massed Compute, use the coupon SECourses. Compare the providers in RunPod vs Massed Compute vs SimplePod.

Related