AndroFlux - V11-3600

Changelog from v8 Configuration

• MAX_NUM_STEPS: 10000 >> 30000

• CHECKPOINTING_STEPS: 100 >> 200

• TRAIN_BATCH_SIZE: 2 (unchanged)

• GRADIENT_ACCUMULATION_STEPS: 1 >> 2

• RESOLUTION: 1024 >> 512

• LEARNING_RATE: 1.0 (unchanged)

• OPTIMIZER: Prodigy (unchanged)

• VALIDATION_GUIDANCE: 3.5 (unchanged)

• VALIDATION_GUIDANCE_REAL: 1.0 >> 2.0

• INSTANCE_PROMPT: Not Set >> "AndroFlux, "

• TRAINER_EXTRA_ARGS:

  • base_model_precision: fp8-quanto (unchanged)

  • lora_rank: 8 >> 16

  • max_grad_norm: 0.01 >> 1.0

  • gradient_precision: Not Set >> unmodified

  • base_model_default_dtype: bf16 >> fp32