feat: add LoRA v2 safetensors and config files

V2 is here! The latest Flux Ghibsky LoRA model has been trained on more images, including faces and synthetic data from V1. While it appears smoother, I have mixed feelings about whether it’s truly an improvement. Training configurations and captioning on the platform that I use can be unpredictable, which may have also influenced the results.

config_v2.yaml

@@ -0,0 +1,60 @@
+job: custom_job
+config:
+  name: flux_train_replicate
+  process:
+  - type: custom_sd_trainer
+    training_folder: output
+    device: cuda:0
+    trigger_word: GHIBSKY
+    network:
+      type: lora
+      linear: 16
+      linear_alpha: 16
+    save:
+      dtype: float16
+      save_every: 1601
+      max_step_saves_to_keep: 1
+    datasets:
+    - folder_path: input_images
+      caption_ext: txt
+      caption_dropout_rate: 0.2
+      shuffle_tokens: false
+      cache_latents_to_disk: false
+      cache_latents: true
+      resolution:
+      - 512
+      - 768
+      - 1024
+    train:
+      batch_size: 5
+      steps: 1600
+      gradient_accumulation_steps: 1
+      train_unet: true
+      train_text_encoder: false
+      content_or_style: balanced
+      gradient_checkpointing: true
+      noise_scheduler: flowmatch
+      optimizer: adamw8bit
+      lr: 0.0004
+      ema_config:
+        use_ema: true
+        ema_decay: 0.99
+      dtype: bf16
+    model:
+      name_or_path: FLUX.1-dev
+      is_flux: true
+      quantize: true
+    sample:
+      sampler: flowmatch
+      sample_every: 1601
+      width: 1024
+      height: 1024
+      prompts: []
+      neg: ''
+      seed: 42
+      walk_seed: true
+      guidance_scale: 3.5
+      sample_steps: 28
+meta:
+  name: flux_train_replicate
+  version: '1.0'

lora_v2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b7a3d6bf24ba01bdc4eda8988f13de91dd94a9af03e15b72e4792c5059ddb84
+size 171969424