Dashboard

LoRA Dataset Prep

Source Images (up to 7)

Drag & drop images here, or browse

JPG, PNG, WEBP — max 20 MB each

Training Settings
This word activates your LoRA when generating. Use something unique.
7 images × 10 = 70 training images
Min Max
Lower = closer to original. Higher = more variation. Keep under 0.6 to preserve identity.
Face Crops
Extracts the top portion of each image as a close-up. Helps the LoRA learn facial detail.
Sets all images at once. Drag the blue line on any thumbnail to fine-tune per-image.
Variation Prompts
One prompt per variation. They cycle — if you have 10 variations and 10 prompts, each gets used once. {trigger} is replaced with your trigger word.
Pipeline Status
  • 1
    Save source imagesWaiting…
  • 2
    Augment via Stable Diffusion Waiting…
  • 3
    Crop, resize & captionWaiting…
  • 4
    Package ZIPWaiting…
What to do with the ZIP — RunPod
  1. Go to RunPod.io → Templates → search Kohya → select kohya_ss GUI → deploy on an RTX 3090 or 4090. Takes ~2 minutes to start.
  2. Upload your dataset — use the RunPod File Uploader (connect to port 2999). Upload lora-dataset.zip to /workspace/.
  3. Unzip it — open Jupyter Lab (port 8888), open a terminal, and run:
    cd /workspace && unzip lora-dataset.zip
    You should now have /workspace/dataset/img/10_{trigger}/.
  4. Open the Kohya GUI on port 3000. Go to the LoRA tab.
  5. Set Pretrained model to any SD 1.5 checkpoint (e.g. runwayml/stable-diffusion-v1-5).
  6. Set Image folder to /workspace/dataset/img. Set Output folder to /workspace/dataset/model. Set Logging folder to /workspace/dataset/log.
  7. Set Trigger word to whatever you typed above. Leave all other settings at defaults. Click Train. Takes ~20–30 minutes on a 3090.
  8. When done, download the .safetensors file from /workspace/dataset/model/ and drop it into /mnt/home/stable-diffusion/models/lora/ on iggy. Use it in A1111 with your trigger word.
  9. Stop your RunPod pod when training finishes — it bills by the hour.