Latest / Big Iron Radio / Homelab Brief - Training LoRAs Locally with Kohya-ss

Transcript

Homelab Brief - Training LoRAs Locally with Kohya-ss · Big Iron Radio

• 0:02
  Welcome to the Homelab Brief from Big Iron.
• 0:06
  Today, training your own image model low R-rays at home with Koya SS.
• 0:12
  Quick note, this is an AI voice drawn from a Big Iron guide.
• 0:16
  Links below.
• 0:18
  First, the lay of the land.
• 0:20
  Koya SS is the de facto toolkit for stable diffusion fine-tuning.
• 0:25
  And your default choice inside it should be a low R-ray, it's a small file.
• 0:29
  It needs only modest Vram, and it mixes with any base model.
• 0:34
  Escalate to low Con or Lycoris only when style and structure come out flat,
• 0:39
  or to full Dreambooth when you genuinely need a multi-gigabyte checkmate and have the Vram to train it.
• 0:45
  Decide the mode first before you touch anything else.
• 0:48
  The guide has a five-question tree subject versus style, your Vram, your file size needs,
• 0:54
  how much fine detail you need, and data set size.
• 0:56
  Pick the wrong mode, and you'll waste hours training the wrong thing.
• 1:01
  Answer those five, then start.
• 1:04
  Here's the part beginners underestimate.
• 1:07
  Data set prep is about 80% of the result.
• 1:10
  Quality over quantity 15 to 30, genuinely great images beats hundreds of mediocre ones.
• 1:17
  And data set prep has real technique.
• 1:20
  Use a consistent captioning scheme a trigger word for a subject, content captions for a style.
• 1:25
  Vary the backgrounds so the model doesn't accidentally learn your couch,
• 1:30
  along with your face, that's called context leakage.
• 1:33
  And use the correct resolution for your base model class.
• 1:37
  On the config knobs, the one that matters most is total steps roughly your image count times the repeats.
• 1:44
  Get the data set in the step count right, and the rest is mostly defaults.
• 1:48
  The full walkthrough with that five-question tree is in the show notes.
• 1:53
  That's your brief.