Melty Style (Melting and Glowing)

This is a melting and glowing style. The prompt should start with melty and can include anything else you can imagine. It will not turn everything into a puddle of slime. Tends to make a lot of spheres and pools of water.

This was trained on a micro dataset of 10 images generated by Flux with a 16:9 aspect ratio suitable for desktop backgrounds. This took about 60 minutes to train and seems to work with a strength from 1.0 - 1.5, so it is probably a little bit under-trained.

This is another test for a different training method, this time using a very small dataset and captions that each have 2 words, melty [something].

Concepts

• melty (melting)

  • planets

  • metal

  • fire

  • stars

  • energy

• glowing (not captioned)

Example Prompts

• melty planets colliding in a stellar explosion

• melty cybernetic diamonds suspended in the mainframe

Training Parameters

params:
  d_coef: 2
  learning_rate: 1.0
  learning_scheduler: cosine_with_restarts
  max_epochs: 15
  network_alpha: 32
  network_dim: 32
  noise_offset: 0.1
  num_cycles: 5
  optimizer_type: prodigy
  optimizer_args: |
    "decouple=True" "weight_decay=0.1" "betas=0.9,0.999" "use_bias_correction=False" "safeguard_warmup=False"
  save_every_n_steps: 40
  sample_every_n_steps: 40
  timestep_sampling: shift
  weight_decay: 0.1

samples: flux_melty_test

and

{
        "datasets": [{
                "resolution": 1280,
                "enable_bucket": true,
                "min_bucket_reso": 768,
                "max_bucket_reso": 1280,
                "bucket_reso_steps": 256,
                "bucket_no_upscale": true,
                "batch_size": 6,
                "subsets": [{
                        "image_dir": "/mnt/training/flux/flux-melty-test",
                        "caption_extension": ".txt",
                        "num_repeats": 5
                }]
        }]
}

This is the same configuration shown in my Flux Nude/NSFW Training Base model, but this LoRA does not include nudity. I wanted to test that configuration without a base LoRA and training from Flux itself. This is the result. Please let me know what you think of it.