[LoRA F.1] crying_with_tears_flux

V2.0

Special thanks:
● 涙/tears (Pony)
● T-ponynai3

The dataset can be downloaded (26 images). The workflow is similar to V1.0, but I changed the generated model, to get anime style training dataset. With additional prompts you can tweak the poses (looking up/down, smiling, etc.).

I found a thing: when I used realistic training data, the lora can generate realistic photos. However, it's hard to generate anime style ones (though add "anime" prompt). On the other hand, using anime training data is good. You can add "photorealistic" prompt when generating to avoid anime style.

(I added the "anime style" tag in training data, maybe it has influence?)

{
  "engine": "kohya",
  "unetLR": 1,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 6,
  "resolution": 512,
  "lrScheduler": "cosine",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 1040,
  "enableBucket": true,
  "networkAlpha": 12,
  "optimizerType": "Prodigy",
  "textEncoderLR": 0,
  "maxTrainEpochs": 20,
  "shuffleCaption": false,
  "trainBatchSize": 3,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

~~~~~~~~~~~~~~~~~~~~~~~o(*￣︶￣*)o~~~~~~~~~~~~~~~~~~~~~~~

示例提示词和参数：
Prompt:  [A girl:snowflakes:5] are falling in the sky, the snowflakes are arranged to show an (abstract pattern of a girl's head:1.2) with girl's face. She is (crying with tears:1.2), photorealistic, masterpiece.
<lora:crying_with_tears_flux_v2.safetensors:1>,

Steps: 30, CFG scale: 3.5, Sampler: Euler, Seed: 303295927047413, width: 1024, height: 832