The base model is the art engine.
SDXL already knows how to draw many kinds of images. It has general knowledge: portraits, animals, flowers, houses, line art, lighting, paper texture, and composition.
SDXL LoRA checkpoint review
Training a small style adapter, one checkpoint at a time.
This page shows what the model is learning as training progresses. It is not the final model. It is a way to see whether the pencil-line style is starting to appear, whether it is too weak, or whether it is getting too rigid.
Plain-English version
What is a LoRA?
The LoRA is a small set of steering notes.
Instead of rebuilding the whole model, LoRA training teaches a compact add-on. When the trigger word is used, that add-on nudges SDXL toward the look learned from the training set.
Checkpoints are progress snapshots.
During training, the system saves checkpoints. Early ones may barely show the style. Later ones may improve, or they may start forcing the same look too strongly. That is why we compare them side by side.
Outputs so far
Checkpoint tabs
Pick one evaluation package, then one checkpoint. Each checkpoint loads its sheet, image set, or settings separately so the page stays readable.
Vote queue
Side-by-side comparisons
Pick the stronger image, tag what changed, and the page will move to the next comparison automatically.
The voting system shows two checkpoint outputs at a time for the same prompt and LoRA strength. Log in as Joy or Krystal, tag each image with anything noticeable, then choose the better image, mark it too close, or mark that neither image is good. Each vote is saved to the shared results dashboard and the next comparison loads automatically; use Admin only for reviewing results and exporting the vote data.
Vote data
Results dashboard
Live totals from Joy and Krystal's review sessions.