MitsuArtXL Style LoRA

A versatile LoRA that maintains high image quality and consistency while playing well with other styles. Recommended weight: 0.6 (full weight can cause random hair color shifts).

Strengths

• Decent hand rendering, particularly with ADetailer

• Strong anatomy consistency across human and non-human subjects

• Handles diverse skin tones well (tested with light, dark, and red skin)

• Works effectively with animal subjects and furry content

• Good weapon rendering, even with complex prompts

• Maintains quality with wide-shot compositions despite slight edge blur

• Compatible with other art styles (tested with Kajin and Naga U)

• Handles chibi style adequately

• Capable of extreme poses while maintaining reasonable anatomy

• Drop shadow effects render properly

Characteristics

• Tends toward bright lighting and vibrant colors

• Best results at 0.6 weight (full weight can cause hair color inconsistency)

• Slight blur effect in wide-shot compositions

Limitations

• Struggles with expressive faces (>_<, DX expressions) unless mixed with other styles

• Some specific poses may be difficult to achieve

• Can have occasional anatomical inconsistencies in certain image areas

• May need style mixing for certain specialized expressions

Technical Notes

• Significantly outperforms epoch 81, particularly in fine detail rendering

• More reliable ADetailer performance compared to epoch 81

• Testing shows consistent performance across various subject matter and compositions

Model Details

SDXL Version:

• Weight: 0.8 - 1.0 (Testing shows 0.6 optimal for epoch 91)

• Trigger: mitsuartstyle

SD 1.5 Version:

• Weight: 0.8 - 1.0

• Trigger: None required

Technical 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

{
  "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 (0.6 for best results with epoch 91)

• Adjust based on your specific needs

• Works well with both simple and complex prompts

• Excellent for style mixing and combinations

📜 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.