RulzTrainer Dataset tools

Dataset tools

The Dataset tab lists every dataset as a card with a few thumbnails. Create dataset makes a folder in datasets/ on the machine running RulzTrainer; Add a folder lists one you already have. Open a dataset to:

• Upload images and .txt / .json captions from the computer your browser runs on to the machine running RulzTrainer (a plain copy when it is the same machine, and the way in on a rented GPU).
• Copy from a folder on the machine itself.
• Edit captions in the grid; they save to the .txt as you type.
• Tick Face ref / Body ref, see the loss watch's verdict on each image, and tick Train on this dataset.

Several datasets can train in one run, each with its own resolutions and repeats (how often each image comes up per epoch), and each can have its own trigger word (needed for LoRA Isolation).

Look consistency filter (face and body outliers)

Finds the pictures that do not look like the rest, after Fizgig's tool.

1. Open the dataset, expand Look consistency filter and choose Face outliers or Body outliers.
2. Tick Baseline on 3 or so pictures that show exactly what you want (a front-facing face; the body you want).
3. Scan. Every picture is compared with the baselines and listed worst first.
   • Face uses the same ArcFace model as the probes: 0.45 and up is a solid match, 0.30 to 0.45 drifting, under 0.30 likely off-look or someone else.
   • Body uses the DINOv2 body window; those scores sit close together, so it reads off the scan's own median and fence. A picture that does not show the neck to mid-thigh gets no score.
4. Auto-suggest marks the pictures under the fence (median minus 1.5 times the spread). Tick or untick others by hand.
5. Move marked out moves them, with their captions, to excluded_by_look/ inside the dataset. That folder is never trained on, nothing is deleted, and Move all back restores everything.

On the Train tab, Warm up look outliers keeps unusual-but-real pictures in the dataset but eases them in: they start at a lower loss weight and reach full weight after a few epochs, while the identity forms from the typical pictures.

LoRA Isolation

A character LoRA tends to put its subject's face on everyone else. Fizgig measured on Krea 2 with two people in one LoRA: 48 of 50 strangers looked like the subject with a normal LoRA, 0 of 50 with Isolation, with the subject's own likeness unchanged. Turn on LoRA Isolation (Train tab) and RulzTrainer adds:

• DOP: every second step each picture also trains with its trigger removed, held to what the model predicts without the LoRA, so without the trigger the LoRA changes nothing.
• A fused trigger: gwelt, a man trains as gweltman. Prompt with the joined word in ComfyUI. Your caption files are never changed.
• Side-by-side pairs: each person gets 16 pictures of their photo next to a generated stranger, so the LoRA learns that the stranger stays a stranger.
• Several subjects: one dataset each, each with its own trigger, balanced so they train equally often. A second person in the LoRA helps even if you only care about one.

It takes about 1.5 to 2 times as long to train. Each dataset row has a Kind: a person gets everything, an animal or thing the fused trigger, a style keeps its trigger as written. Measured by Fizgig; not yet measured on RulzTrainer's own runs.