I made it via a lora trainer on hugginface, specifically this ones .. <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/autotrain-projects/train-flux-lora-ease">https://huggingface.co/spaces/autotrain-projects/train-flux-lora-ease</a>

Phoebe Combes

<a target="_blank" rel="ugc" href="https://huggingface.co/black-forest-labs/FLUX.1-schnell">black-forest-labs/FLUX.1-schnell</a>version 1.0 , just made it from a sett of pictures of her from listal device: cuda:0model: is_flux: true quantize: truenetwork: 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: lorasample: 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: 1024save: dtype: float16 hf_private: true max_step_saves_to_keep: 4 push_to_hub: true save_every: 10000train: 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

Files with SHA256: 567c8f7ddcd030e06693c7e4cb7d3f64557fae745fa2e07096f389ddee1c565e