Training Settings Comparison

05_training_config_female_actor.yaml vs 06_training_config_male_actor.yaml

Female actor mzyeoja (v3) Male actor joseonnamja (v3) 3 differences
~60
Total settings
57
Identical
3
Different

Key Differences

trigger_word
mzyeoja
joseonnamja
num_repeats
2
1
Sample prompt
mzyeoja woman, ...
joseonnamja man, ...

Full YAML Diff

Female actor mzyeoja_v3
Male actor joseonnamja_v3
Base config
Base config
job: extension
job: extension
name: mzyeoja_v3
name: joseonnamja_v3
trigger_word: mzyeoja
trigger_word: joseonnamja
Network
Network
type: lora
type: lora
linear: 16 / linear_alpha: 16
linear: 16 / linear_alpha: 16
conv: 8 / conv_alpha: 8
conv: 8 / conv_alpha: 8
Saving
Saving
dtype: bf16
dtype: bf16
save_every: 100
save_every: 100
save_format: diffusers
save_format: diffusers
Dataset
Dataset
folder_path: datasets\mzyeoja
folder_path: datasets\joseonnamja
caption_ext: txt
caption_ext: txt
caption_dropout_rate: 0.05
caption_dropout_rate: 0.05
cache_latents_to_disk: false
cache_latents_to_disk: false
resolution: [1024, 768]
resolution: [1024, 768]
num_repeats: 2
num_repeats: 1
Training parameters
Training parameters
batch_size: 1
batch_size: 1
steps: 3000
steps: 3000
gradient_accumulation: 2
gradient_accumulation: 2
optimizer: adamw
optimizer: adamw
lr: 8.0e-05
lr: 8.0e-05
dtype: bf16
dtype: bf16
ema_decay: 0.99
ema_decay: 0.99
noise_scheduler: flowmatch
noise_scheduler: flowmatch
loss_type: mse
loss_type: mse
Model
Model
name_or_path: FLUX.2-klein-base-9B
name_or_path: FLUX.2-klein-base-9B
quantize: true / qtype: qfloat8
quantize: true / qtype: qfloat8
arch: flux2_klein_9b
arch: flux2_klein_9b
Sampling
Sampling
sample_every: 100 / guidance_scale: 4
sample_every: 100 / guidance_scale: 4
sample_steps: 50 / walk_seed: true
sample_steps: 50 / walk_seed: true

Sample Prompt Comparison

Female actor (mzyeoja) Male actor (joseonnamja)
mzyeoja woman, chest up portrait, 85mm photo, neutral expression, indoor soft light seed: 42 / mult: 1.0
joseonnamja man, chest up portrait, 85mm photo, neutral expression, indoor soft light seed: 42 / mult: 1.0
mzyeoja woman, three-quarter view, smiling, outdoor daylight, upper body seed: 43 / mult: 1.0
joseonnamja man, three-quarter view, smiling, outdoor daylight, upper body seed: 43 / mult: 1.0
mzyeoja woman, portrait photo, serious expression, wearing dark jacket, upper body seed: 44 / mult: 1.0
joseonnamja man, portrait photo, serious expression, wearing dark jacket, upper body seed: 44 / mult: 1.0
Every prompt is generated as a 1.0 + 0.8 multiplier pair (6 images total). When the 0.8 version starts looking better, you've crossed into overfit.

Why num_repeats Differs

Female actor (num_repeats: 2)
Only 12 images — relatively small.
12 images x 2 repeat = 24 exposure/step
Comparable exposure to djdante's 29 x 1 = 29.
Male actor (num_repeats: 1)
16 images — sufficient.
16 images x 1 repeat = 16 exposure/step
Strategy: maximize variety.