Phoebe Combes版本v1.0 (ID: 873014)

I made it via a lora trainer on hugginface, specifically this ones ..

https://huggingface.co/spaces/autotrain-projects/train-flux-lora-ease

描述:

black-forest-labs/FLUX.1-schnell

version 1.0 , just made it from a sett of pictures of her from listal

device: cuda:0

model:

is_flux: true

quantize: true

network:

linear: 16 #it will overcome the 'rank' parameter

linear_alpha: 16 #you can have an alpha different than the ranking if you'd like

type: lora

sample:

guidance_scale: 3.5

height: 1024

neg: '' #doesn't work for FLUX

sample_every: 1000

sample_steps: 28

sampler: flowmatch

seed: 42

walk_seed: true

width: 1024

save:

dtype: float16

hf_private: true

max_step_saves_to_keep: 4

push_to_hub: true

save_every: 10000

train:

batch_size: 1

dtype: bf16

ema_config:

ema_decay: 0.99

use_ema: true

gradient_accumulation_steps: 1

gradient_checkpointing: true

noise_scheduler: flowmatch

optimizer: adamw8bit #options: prodigy, dadaptation, adamw, adamw8bit, lion, lion8bit

train_text_encoder: false #probably doesn't work for flux

train_unet: true

训练词语: Phoebe