🧊Guidance Samplers

Static thresholding: We refer to elementwise clipping the x-prediction to [−1, 1] as static thresholding. This method was in fact used but not emphasized in previous work, and to our knowledge, its importance has not been investigated in the context of guided sampling. We discover that static thresholding is essential to sampling with large guidance weights and prevents generating blank images. Nonetheless, static thresholding still results in over-saturated and less detailed images as the guidance weight further increases

The images created by FonAI with the description "a dog swimming in the sky"

Dynamic thresholding: We introduce a new dynamic thresholding method: at each sampling step we setto a certain percentile absolute pixel value in , and if, then we thresholdto the rangeand then divide by. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing pixels from saturation at each step. We find that dynamic thresholding results in significantly better photorealism as well as better image-text alignment, especially when using very large guidance weights. This sampler is suitable for art schools like NFTs, digital artwork, vector design, etc...

The images created by FonAI with the description "a big dog swimming in the sky with gradient clouds and schools of colorful fish swimming around, digital art"