THE TRAINING PROCESS
--------------------------------------------------------------------------------
We trained this beast through approximately 40 epochs / 9,400 steps using a
dataset of videos, each running 6-8 seconds. Two H200s working overtime for
about 20 hours straight. Why so long? Because we're not amateurs. We chose a
more stable, longer training approach after our attempts at faster, more
aggressive learning blew up the gradients. You can't rush perfection.
HYPERPARAMETERS
--------------------------------------------------------------------------------
LORA_RANK=32
MAX_EPOCHS=200
LEARNING_RATE=3e-4
RESOLUTION_LIST="512, 512"
TARGET_FRAMES="1, 57, 117"
FRAME_EXTRACTION="head"
NUM_REPEATS=5
USAGE & SETTINGS
--------------------------------------------------------------------------------
Works beautifully at a straight 1.0 strength for both high noise and low noise.
But here's the thing — if you want to play it safe and reduce the risk of
anatomical distortion…
