BikeMaker版本SD v2.2 1024 (ID: 400073)

BikeMaker v2

• SDXL version weight 0.1-1

• reworked with more design features and details

• SD 1.5 version weight 0.1-1

• 1024 version fixes higher resolution generations in SD:

BikeMaker v1

SDXL (1024x1024) note: Use also negative weights, check examples. Choose versions from the menu on top.

BikeMaker is a tool for generating all types of—you guessed it—bikes. It was trained on an in-house developed dataset of 180 designs with interesting concept features. The model either fixes the input or makes it more interesting for variations.

Use Hires. fix (4x_RealisticRescaler, 4x-Ultrasharp upscalers with denoising strength around 0.3-0.6) for best details in txt2img!

You can combine it with my (or any other) models for added (sur)realism. Check the gallery for examples.

Experiment with weights around 0.1-1 (lower values will tend to create a modern bike or electro bike, higher sci-fi motorbikes), CLIP SKIP 1-2 (2 works better for some stylized models). The model works with a bike token (try to use a combination bike|scooter, vehicle|bike, and so on). It will also style a combined vehicle.

LyCORIS/LoHA experiment trained 768x768, NR 32 + NA 16 dimensions. I am not aiming for technical realism with this series, the models are more focused on design ideas. If you want to learn more about AI-augmented design and algorithmic art, check my web pages.

For my notes on training and use of stable diffusion models, check https://sandner.art/articles/ai/

Post some images of what you can create with the model here, it helps to make future versions.

If you feel generous, you may also throw some buzz in this direction or support my research on Ko-fi https://ko-fi.com/airesearch.

Have fun rendering!

描述:

{

  "unetLR": 0.0005,

  "clipSkip": 1,

  "loraType": "lora",

  "keepTokens": 0,

  "networkDim": 32,

  "numRepeats": 2,

  "resolution": 1024,

  "lrScheduler": "cosine_with_restarts",

  "minSnrGamma": 5,

  "noiseOffset": 0.1,

  "targetSteps": 6120,

  "enableBucket": true,

  "networkAlpha": 16,

  "optimizerType": "AdamW8Bit",

  "textEncoderLR": 0.00005,

  "maxTrainEpochs": 8,

  "shuffleCaption": false,

  "trainBatchSize": 2,

  "flipAugmentation": true,

  "lrSchedulerNumCycles": 3

}

训练词语: