training_config.yaml

---
job: extension
config:
  # this name will be the folder and filename name
  name: "supernal"
  process:
    - type: 'sd_trainer'
      # root folder to save training sessions/samples/weights
      training_folder: "/opt/stationthis/jobs/training-1774631660875/output"
      # uncomment to see performance stats in the terminal every N steps
#      performance_log_every: 1000
      device: cuda:0
      # if a trigger word is specified, it will be added to captions of training data if it does not already exist
      # alternatively, in your captions you can add [trigger] and it will be replaced with the trigger word
      trigger_word: "supernal"
      network:
        type: "lora"
        linear: 32
        linear_alpha: 32
      save:
        dtype: float16 # precision to save
        save_every: 250 # save every this many steps (min of 250 or total steps)
        max_step_saves_to_keep: 2 # how many intermittent saves to keep
        push_to_hub: false #change this to True to push your trained model to Hugging Face.
        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
#       hf_repo_id: your-username/your-model-slug
#       hf_private: true #whether the repo is private or public
      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: "/opt/stationthis/jobs/training-1774631660875/dataset"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          shuffle_tokens: false  # shuffle caption order, split by commas
          cache_latents_to_disk: true  # leave this true unless you know what you're doing
          resolution: [ 512, 768, 1024 ]  # flux enjoys multiple resolutions
      train:
        batch_size: 1
        steps: 4000  # total number of steps to train 500 - 4000 is a good range
        gradient_accumulation_steps: 1
        train_unet: true
        train_text_encoder: false  # probably won't work with flux
        gradient_checkpointing: true  # need the on unless you have a ton of vram
        noise_scheduler: "flowmatch" # for training only
        optimizer: "adamw8bit"
        lr: 1e-4
        # Skip baseline samples, only sample during/after training
        skip_first_sample: true
        # set to true to completely disable sampling
        # disable_sampling: true
        # uncomment to use new vell curved weighting. Experimental but may produce better results
#        linear_timesteps: true

        # ema will smooth out learning, but could slow it down. Recommended to leave on.
        ema_config:
          use_ema: true
          ema_decay: 0.99

        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
        dtype: bf16
      model:
        # huggingface model name or path
        name_or_path: "black-forest-labs/FLUX.1-dev"
        is_flux: true
        quantize: true  # run 8bit mixed precision
#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
      sample:
        sampler: "flowmatch" # must match train.noise_scheduler
        sample_every: 4000 # sample at final step (TRAIN_STEPS - 1 for 0-indexed)
        width: 1024
        height: 1024
        prompts:
          # Dynamic prompts from dataset captions (injected by launch-training.js)
          # These will be replaced with actual captions for HuggingFace samples
          - "supernal, This is a digital drawing in a manga style, featuring a large, blue and yellow robotic creature with shark-like features, including a pointed snout and red eyes. The creature's body is segmented and mechanical, with visible gears and cables. The background showcases a detailed, intricate mechanical structure, filled with gears and pipes, resembling a futuristic spaceship or industrial complex. The left side of the image has a vertical black and pink text block with stylized letters, and a smaller black and white inset at the bottom left corner. The overall color palette includes shades of blue, yellow, red, and gray."
          - "supernal, This is a digitally-created, hyper-realistic close-up photograph of a human face with a fantastical twist. The central focus is a striking green eye with detailed iris and long, dark eyelashes. The left side of the face is adorned with a textured, crystalline pattern in shades of white, teal, and black, resembling ice or gemstones. The right side of the face is mostly plain. The background is black with white, futuristic text on the left and right edges. The text on the right is in a vertical column, and the overall style is reminiscent of sci-fi or fantasy art."
          - "supernal, This is a highly detailed, digital illustration in a manga/anime style. The central figure is a close-up of a white-skinned female with her head tilted back, eyes closed, and lips slightly parted. Her hair is intricately adorned with a complex, mechanical-like headdress featuring gears, wires, and metallic textures. The background is a vibrant, colorful outer space scene with a bright red and blue planet. The left side of the image has vertical text in bold, white letters. The right side features a dense block of Japanese text. The overall color palette is stark, with high contrast between the black, white, and colorful background elements."
          - "supernal, This digital artwork depicts a towering, armored warrior statue standing amidst fluffy white clouds against a vibrant blue sky. The statue, clad in intricately detailed, silver metallic armor with elaborate patterns, holds a large, ornate sword in its right hand. Its helmet features a prominent, pointed crest, and a halo-like ring hovers above its head. The sun, depicted as a bright, glowing red circle, radiates beams of light from the upper left corner, casting a dramatic effect. Red vertical text runs along the right side of the image. The overall artistic technique blends fantasy and surrealism, with a strong emphasis on metallic textures and celestial elements."
        neg: ""  # not used on flux
        seed: 42
        walk_seed: true
        guidance_scale: 4
        sample_steps: 20
# you can add any additional meta info here. [name] is replaced with config name at top
meta:
  name: "supernal"
  version: '1.0'