Dataset:
Caption format:
v1: heterochromia, orange eye, ice blue eye — no directional control, base trigger only
v2: heterochromia, heterochromia_iridum, left_eye_orange, right_eye_blue, different_colored_eyes — full directional control, multiple triggers
Key improvements:
Added heterochromia_iridum as secondary trigger
Added different_colored_eyes as tertiary trigger
left_eye_X and right_eye_X teach the model which color goes on which side
Red sclera format cleaned up and standardized
Underscores in all trigger words for cleaner token separation
Training settings:
v1: 15 repeats, 15 epochs = 31,950 steps, ~17 hours
v2: 10 repeats, 10 epochs = 5,250 steps, ~3.5 hours
v2 loss finished at 0.109 vs v1's 0.117 — slightly better convergence
What v2 should do better:
Consistent left/right eye color assignment when prompted
Stronger response to specific color combinations
More reliable red sclera activation
Better generalization due to cleaner captions
v2 LoRA trigger words:
Main triggers (always include):
heterochromia, heterochromia_iridum, different_colored_eyes
Directional eye control:
left_eye_[color], right_eye_[color]
Color options used in training:
violet, cyan, silver, green, amber, emerald, sapphire,
turquoise, teal, pink, burgundy, rose, yellow, purple,
white, ruby, red, magenta, lime, jade, grey, gold, crimson, black
Red sclera format:
left_eye_red_sclera_[iris color]_iris
right_eye_red_sclera_[iris color]_iris
Full example prompts:
heterochromia, heterochromia_iridum, left_eye_blue, right_eye_red, different_colored_eyes, 1girl, white hair
heterochromia, heterochromia_iridum, left_eye_red_sclera_silver_iris, right_eye_violet, different_colored_eyes, 1girl, silver hair, gothic outfit
heterochromia, heterochromia_iridum, left_eye_amber, right_eye_cyan, different_colored_eyes, 1girl
Start with LoRA weight 0.7-0.8 and adjust from there.
