Best Free LoRA Training Tools in 2026: Local and Cloud Options Compared

Training a LoRA involves three separate problems: preparing your images, tagging your dataset, and running the actual training. Most lists treat this as one problem. It is not.

The tool that trains your model is not the tool that tags your dataset. The tool that resizes your images is not the tool that writes your captions. Each stage has its own options with its own trade-offs.

Here is what is actually worth using in 2026, split by stage.

---

Stage 1: Image Preparation

Before you train anything, your images need to be consistent in resolution and composition. Kohya_ss and OneTrainer both expect images at a fixed resolution, square or bucketed.

Birme (Free, Browser)

Birme is a batch image resizer that runs in the browser. Drag in a folder, set your target resolution, and download. For SD 1.5 LoRAs use 512x512. For SDXL and Flux, use 1024x1024.

Fast and no-install. Where it falls short: no smart cropping. If your subject is not centered, you will crop wrong and need to fix it manually.

ImageMagick (Free, CLI)

For larger datasets or when you want this step in a script, ImageMagick handles batch crop and resize in one command:

magick mogrify -resize 512x512^ -gravity center -extent 512x512 *.jpg

This center-crops every image to exact square. Works well for character datasets where the subject is usually centered.

---

Stage 2: Dataset Tagging

Tagging is where most people spend the most time and where bad choices show up most clearly in output quality. Under-tagging causes attribute bleeding. Over-tagging confuses the model about what it is actually supposed to learn.

Free Dataset Tagger (Browser, No Install)

I built a free dataset tagger that runs entirely in your browser using the WD14 model. Upload images, get Danbooru-convention tags, edit them inline, and export .txt files ready for training.

No Python environment, no GPU, no setup. Works on any device.

WD14 Tagger via A1111 or Kohya (Local, GPU)

WD ViT Tagger v3 by SmilingWolf is the standard automatic tagger in the community. It is built into Kohya_ss's captioning tool and available as an AUTOMATIC1111 extension.

For large datasets (100+ images), running WD14 locally is faster than a browser tool. But you still need to review and edit the output, especially for character-specific attributes the model will not pick up automatically.

BooruDatasetTagManager (Windows, Free)

BooruDatasetTagManager is a GUI for bulk-editing caption files. After a WD14 first pass, use this to add your trigger word across every file, fix wrong tags in bulk, and remove tags for attributes you do not want the LoRA to learn.

---

Stage 3: Training

Kohya_ss (Free, Local)

Kohya_ss is the most widely used local LoRA trainer. It supports SD 1.5, SDXL, and Flux, has a web UI, and gives you full control over every training parameter including learning rate, network dimensions, optimizer, and noise offset.

Initial setup takes time. The documentation is scattered across GitHub issues and the community Discord. But once running, it is the most capable free option available.

Minimum VRAM: 8GB for SD 1.5, 12-16GB for SDXL, 16-24GB for Flux.

OneTrainer (Free, Local)

OneTrainer is the cleaner alternative. The UI is more organized, defaults are better for SDXL and Flux, and you do not need to know every parameter to get decent results out of the box.

If Kohya_ss feels overwhelming or you are training SDXL/Flux for the first time, start here.

Civitai On-Site Trainer (Free Tier, Cloud)

Civitai's LoRA trainer lets you train directly on their platform with no local setup. Free tier available with limits on steps and dataset size. Good for testing a concept without a local GPU commitment.

Downside: less control than local training, and free tier jobs queue during high traffic.

PixelDojo and BasedLabs (Free Tier, Cloud)

PixelDojo and BasedLabs both offer cloud Flux LoRA training with free tiers. Fast and beginner-friendly, but you give up control over training parameters. Fine for quick character or style LoRAs.

---

Recommended Setup by Situation

No GPU, just starting out: Birme for prep, browser dataset tagger for captions, Civitai or PixelDojo for training.

Have a GPU (8-12GB VRAM): Birme for prep, WD14 via Kohya + BooruDatasetTagManager for captions, OneTrainer for training.

Experienced, large dataset, full control: ImageMagick for prep, WD14 + custom scripts for captions, Kohya_ss for training.

---

What Actually Makes the Difference

Tools execute the process you design. They do not fix a bad dataset.

Twenty clean, consistently tagged images with a unique trigger word will produce a better LoRA than two hundred noisy images with inconsistent captions. Spend more time on your dataset than on picking a trainer. The feedback loop in LoRA training is short, so iterate fast and fix the dataset when the output is wrong.