I'm going to try that

Done: https://civitai.com/models/2449377?modelVersionId=2754120

https://github.com/kohya-ss/musubi-tuner

Parameters of flux_2_train_network.py:

--fp8_base --fp8_scaled --fp8_text_encoder --vae_dtype bfloat16 \

--sdpa --mixed_precision bf16 --gradient_checkpointing

--timestep_sampling flux2_shift --weighting_scheme none

--optimizer_type adamw8bit --learning_rate 2e-4 --lr_scheduler cosine

--max_data_loader_n_workers 2 --persistent_data_loader_workers

--network_module networks.lora_flux_2 --network_dim 32 --network_alpha 32

Dataset:

[general]

caption_extension = ".txt"

batch_size = 1

enable_bucket = true

bucket_no_upscale = false

[[datasets]]

image_directory = "images"

cache_directory = "c4ch3d1r_512"

resolution = [512, 512]

num_repeats = 2

[[datasets]]

image_directory = "images"

cache_directory = "c4ch3d1r_768"

resolution = [768, 768]

num_repeats = 1