Background
There are two models we care about. The core model operates on a latent map of defined shape and returns another latent map (denoised) and a low level resolution grid. The decoder model, which takes in the diffusion models latent map output and returns a full grid. Both of these are models have different shapes input / output shape.
Description
Use placeholders for each value for the core model and decoder separately:
- Input shape: height, width, and channel count
- Output shape: the same, for every output the model returns
- Value range : what band the numbers are expected to be in
- Precision: float32 or float16
Then create an object that holds these shapes, acting as the component boundary between model input / output. Real output shapes can be found from the paper or hugging face in a config.json file.
Background
There are two models we care about. The core model operates on a latent map of defined shape and returns another latent map (denoised) and a low level resolution grid. The decoder model, which takes in the diffusion models latent map output and returns a full grid. Both of these are models have different shapes input / output shape.
Description
Use placeholders for each value for the core model and decoder separately:
Then create an object that holds these shapes, acting as the component boundary between model input / output. Real output shapes can be found from the paper or hugging face in a
config.jsonfile.