train_flex_redux.yaml

---
job: extension
config:
  # this name will be the folder and filename name
  name: "my_first_flex_redux_finetune_v1"
  process:
    - type: 'sd_trainer'
      # root folder to save training sessions/samples/weights
      training_folder: "output"
      # uncomment to see performance stats in the terminal every N steps
#      performance_log_every: 1000
      device: cuda:0
      adapter:
        type: "redux"
        # you can finetune an existing adapter or start from scratch. Set to null to start from scratch
        name_or_path: '/local/path/to/redux_adapter_to_finetune.safetensors'
        # name_or_path: null
        # image_encoder_path: 'google/siglip-so400m-patch14-384' # Flux.1 redux adapter
        image_encoder_path: 'google/siglip2-so400m-patch16-512' # Flex.1 512 redux adapter
        # image_encoder_arch: 'siglip' # for Flux.1
        image_encoder_arch: 'siglip2'
        # You need a control input for each sample. Best to do squares for both images
        test_img_path: 
          - "/path/to/x_01.jpg"
          - "/path/to/x_02.jpg"
          - "/path/to/x_03.jpg"
          - "/path/to/x_04.jpg"
          - "/path/to/x_05.jpg"
          - "/path/to/x_06.jpg"
          - "/path/to/x_07.jpg"
          - "/path/to/x_08.jpg"
          - "/path/to/x_09.jpg"
          - "/path/to/x_10.jpg"
        clip_layer: 'last_hidden_state'
        train: true
      save:
        dtype: bf16 # precision to save
        save_every: 250 # save every this many steps
        max_step_saves_to_keep: 4 
      datasets:
        # datasets are a folder of images. captions need to be txt files with the same name as the image
        # for instance image2.jpg and image2.txt. Only jpg, jpeg, and png are supported currently
        # images will automatically be resized and bucketed into the resolution specified
        # on windows, escape back slashes with another backslash so
        # "C:\\path\\to\\images\\folder"
        - folder_path: "/path/to/images/folder"
          # clip_image_path is directory containting your control images. They must have filename as their train image. (extension does not matter)
          # for normal redux, we are just recreating the same image, so you can use the same folder path above
          clip_image_path: "/path/to/control/images/folder"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          resolution: [ 512, 768, 1024 ]  # flex enjoys multiple resolutions
      train:
        # this is what I used for the 24GB card, but feel free to adjust
        # total batch size is 6 here
        batch_size: 3
        gradient_accumulation: 2

        # captions are not needed for this training, we cache a blank proompt and rely on the vision encoder
        unload_text_encoder: true

        loss_type: "mse"
        train_unet: true
        train_text_encoder: false
        steps: 4000000  # I set this very high and stop when I like the results
        content_or_style: balanced  # content, style, balanced
        gradient_checkpointing: true
        noise_scheduler: "flowmatch" # or "ddpm", "lms", "euler_a"
        timestep_type: "flux_shift" 
        optimizer: "adamw8bit"
        lr: 1e-4

        # this is for Flex.1, comment this out for FLUX.1-dev
        bypass_guidance_embedding: true

        dtype: bf16
        ema_config:
          use_ema: true
          ema_decay: 0.99
      model:
        name_or_path: "ostris/Flex.1-alpha"
        is_flux: true
        quantize: true
        text_encoder_bits: 8
      sample:
        sampler: "flowmatch" # must match train.noise_scheduler
        sample_every: 250 # sample every this many steps
        sample_start_step: 0 # start sampling at this step
        width: 1024
        height: 1024
        # I leave half blank to test prompt and unprompted
        prompts:
          - "woman with red hair, playing chess at the park, bomb going off in the background"
          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
          - "a bear building a log cabin in the snow covered mountains"
          - ""
          - ""
          - ""
          - ""
          - ""
        neg: ""
        seed: 42
        walk_seed: true
        guidance_scale: 4
        sample_steps: 25
        network_multiplier: 1.0

# you can add any additional meta info here. [name] is replaced with config name at top
meta:
  name: "[name]"
  version: '1.0'