config.generation.sampling_mode |
str |
'posterior' |
Latent sampling source: 'posterior' or 'prior'. |
config.generation.posterior_skip_source |
str |
'original' |
Encoder skip source in posterior mode: 'original' uses the input image; 'transformed' uses an image paired with the generated target mask (conditional ConvNeXt only). |
config.generation.variation_strength |
float |
0.5 |
Strength of random variation during generation; non-negative. |
config.generation.clamp_output |
bool |
False |
Clamp generated image values to [0, 1]. |
config.generation.background_threshold |
float |
0.01 |
Relative threshold for deriving a mask from generated images. |
config.generation.variants_per_real_anomaly |
int |
3 |
Number of synthetic variants per real anomaly; positive. |
config.generation.feedback.enabled |
bool |
False |
Enable image similarity feedback and repeated generation attempts. |
config.generation.feedback.similarity_threshold |
float |
0.8 |
Minimum similarity for accepting a variant; range [0, 1]. |
config.generation.feedback.threshold_relaxation_factor |
float |
0.9 |
Factor for relaxing the threshold; range (0, 1]. |
config.generation.feedback.max_attempts |
int |
100 |
Maximum number of generation attempts; positive. |