MitsuArtXL Style LoRA

A carefully crafted LoRA trained on MitsuArt's distinctive artistic style. This model aims to capture the unique characteristics and aesthetics of MitsuArt's work, offering creators a powerful tool for generating similar artistic outputs.

---

? Usage Guide

SDXL Version:

• Weight: 0.8 - 1.0

• Trigger: mitsuartstyle

SD 1.5 Version:

• Weight: 0.8 - 1.0

• Trigger: None required

---

⚙️ Training Details

Base Resolution: 1024 Network Dim: 32 Network Alpha: 16 Optimizer: Adafactor Scheduler: Cosine with Restarts Training Steps: 2525 Epochs: 100 Batch Size: 8

? Training Configuration

{

  "engine": "kohya",

  "unetLR": 0.0001,

  "clipSkip": 1,

  "loraType": "lora",

  "keepTokens": 1,

  "networkDim": 32,

  "numRepeats": 1,

  "resolution": 1024,

  "lrScheduler": "cosine_with_restarts",

  "minSnrGamma": 5,

  "noiseOffset": 0.1,

  "targetSteps": 2525,

  "enableBucket": true,

  "networkAlpha": 16,

  "optimizerType": "Adafactor",

  "textEncoderLR": 0.00005,

  "maxTrainEpochs": 100,

  "shuffleCaption": true,

  "trainBatchSize": 8,

  "flipAugmentation": true,

  "lrSchedulerNumCycles": 3

}

? Tips

• Start with recommended weights

• Adjust based on your specific needs

• Works well with both simple and complex prompts

? Credit

Based on the artistic style of MitsuArt. Please respect the original artist's work when using this LoRA.

---

If you find this LoRA useful, don't forget to leave a like and share your creations! Feedback and suggestions are always welcome.

描述:

Trained LoRA to capture Mitsuart's art style on SDXL. Built with 134 curated images over 100 epochs.

Works best at 1.0 strength, but can use 0.6 - 0.8 for subtle influence. Trigger: "mitsuartstyle"

Training:

- 134 images × 100 epochs

- Dim/Conv: 32, Alpha: 16

- AdaFactor+ Huber loss

- Cosine w/restarts scheduler

- LR: 0.0001 + Unet LR: 0.00005

- Dropouts: 0.08

- Base: NoobAI Epsilon 1.0

Enjoy! ?

训练词语: mitsuartstyle