it's_a_test版本f4 (ID: 885974)

these models shouldn't be used, it's about a private testing of different training parameters.

I can't prevent you from downloading it =(

Maybe in the future I will change this to a public space with more info to help people to train Flux D

描述:

accelerate launch "./sd-scripts/flux_train_network.py" ^

--pretrained_model_name_or_path "D:\kohya_ss\flux\unet\flux1DevFp8_v10.safetensors" ^

--clip_l "D:/stable-diffusion-webui-forge-main/models/text_encoder/flux/clip_l.safetensors" ^

--t5xxl "D:/stable-diffusion-webui-forge-main/models/text_encoder/flux/flux1T5TextEncoder_stock.safetensors" ^

--ae "D:/stable-diffusion-webui-forge-main/models/VAE/flux/flux1DevVAE_stock.safetensors" ^

--train_data_dir "D:\kohya_ss\flux\img" ^

--output_dir "D:/kohya_ss/flux/models/whimsical_flux4" ^

--output_name "whimsical_flux4" ^

--save_model_as safetensors ^

--sdpa ^

--persistent_data_loader_workers ^

--max_data_loader_n_workers 2 ^

--seed 777 ^

--gradient_checkpointing ^

--mixed_precision bf16 ^

--save_precision fp16 ^

--network_module networks.lora_flux ^

--network_dim 24 ^

--network_alpha 24 ^

--network_args train_blocks=single rank_dropout=0.1 module_dropout=0.1 ^

--optimizer_type lion ^

--lr_scheduler cosine ^

--learning_rate 0.001 ^

--network_train_unet_only ^

--timestep_sampling sigmoid ^

--model_prediction_type raw ^

--guidance_scale 1.0 ^

--loss_type l2 ^

--resolution 512,512 ^

--cache_text_encoder_outputs ^

--cache_text_encoder_outputs_to_disk ^

--fp8_base ^

--enable_bucket ^

--bucket_no_upscale ^

--split_mode ^

--prior_loss_weight 1.0 ^

--caption_extension .txt ^

--keep_tokens 0 ^

--network_dropout 0.1 ^

--max_grad_norm 1.0 ^

--min_snr_gamma 10 ^

--save_every_n_epochs 1 ^

--train_batch_size "2" ^

--multires_noise_iterations 6 ^

--multires_noise_discount 0.3 ^

--optimizer_args weight_decay=0.4 betas="[0.95, 0.98]" ^

--max_train_epochs 30 ^

--enable_wildcard

训练词语: