Generate
Describe what you want. The model is chosen for you unless you pick one.
How this works
Good prompts
- Describe the scene, not the person: lighting, setting, clothing, pose, camera framing.
- For video, describe motion and the camera — she turns her head, camera pushes in slowly. The still already supplies the subject.
- Pick a character to put a trained face in the shot. The trigger word is added for you.
- Leave the model on Auto unless you have a reason. It tells you what it chose and why before you press Generate.
Common mistakes
- Describing the face when a character is selected — it fights the LoRA. Say portrait, soft light, not brown eyes, round face.
- Expecting a character in Flux or SD3.5. Face LoRAs are SDXL only, so choosing a character forces an SDXL model.
- Very long negatives. Most of these models do better with a short one, or none at all.
- Generating while training runs. The GPU cannot do both; it will refuse until training finishes.
Video takes minutes, not seconds. Start at 33 frames to check the
motion is right before going longer. Frame count must be 4n+1 —
33, 49, 65, 81.
Add a character
One LoRA per person. Lowercase letters, digits, underscore or hyphen.
Train a LoRA
Training uses the whole GPU, so generating is unavailable while it runs. Expect 20–30 minutes.
Character
—
Step
—
Loss
—
Elapsed
—
Live log
Not running.
Before you train: the photos
This matters more than any setting on this page. A good set trains well on defaults; a poor set cannot be rescued by tuning.
25–40
photos of one person
1024px+
on the short side
varied
light, angle, clothing
Do this
- Mix the framing. Around two thirds head and shoulders, a quarter waist-up, the rest full body, plus two or three close-ups.
- Vary the light. Daylight, indoors, shade, evening.
- Vary the angle. Straight on, and three-quarter from both sides.
- Change the clothes and the background between shots.
- Use the originals from the camera or phone gallery.
- Keep them sharp and in focus.
Avoid
- Screenshots. Too small and already re-compressed. This is the most common reason a LoRA comes out soft.
- Near-identical shots from the same moment.
- Other people in frame — it cannot tell whose face to learn.
- Sunglasses or anything covering the face.
- Beauty filters and heavy retouching.
- AI-upscaled images — they teach plasticky skin.
- Cut-out backgrounds. Real varied backgrounds look better.
Anything present in every photo gets absorbed into the trigger word.
Thirty photos in the same red jacket means the trigger learns the jacket, and
you will not be able to prompt it away. Variety is what keeps the LoRA
controllable.
What happens when you train
- Prepare — duplicates and blurry shots are dropped, images resized, and EXIF including GPS is stripped. Nothing leaves this machine.
- Caption — each photo gets tags describing what varies: pose, clothing, lighting, background. Identity tags are deliberately removed, so the face binds to your trigger word instead of to the word “brown eyes”.
- Train — roughly 1800 steps, about 20–30 minutes, saving one checkpoint per epoch.
- Choose an epoch in the Characters tab, then generate with it.
Choosing an epoch
Training saves the same LoRA at twelve increasing strengths. You use exactly one. More training is not better.
Early (1–3)Weak likeness. Follows your prompt well.
Middle (4–8)Likeness formed, still flexible. Usually the one to pick.
Late (9–12)Strongest likeness, but starts ignoring your prompt.
If a style prompt like watercolour painting still returns a photo, the
epoch is too strong. Drop to an earlier one, or lower the LoRA weight to
around 0.6.
Security
Two-factor authentication. Strongly recommended before this is reachable from outside your home network.
Checking…
Scan this with Google Authenticator, Authy or 1Password, then enter the 6-digit code it shows.
Cannot scan? Enter this key by hand:
Save these recovery codes now. Each works once, and this is the
only time they are shown. Without them, losing your phone means losing
access to this machine.
Two-factor is on.