This is a fast, minimal port of Boris Dayma's [DALL·E Mini](https://github.com/borisdayma/dalle-mini) (with mega weights). It has been stripped down for inference and converted to PyTorch. The only third party dependencies are numpy, requests, pillow and torch.
Here's a more detailed breakdown of performance on an A100. Credit to [@technobird22](https://github.com/technobird22) and his [NeoGen](https://github.com/technobird22/NeoGen) discord bot for the graph.
The required models will be downloaded to `models_root` if they are not already there. Set the `dtype` to `torch.float16` to save GPU memory. If you have an Ampere architecture GPU you can use `torch.bfloat16`. Set the `device` to either "cuda" or "cpu". Once everything has finished initializing, call `generate_image` with some text as many times as you want. Use a positive `seed` for reproducible results. Higher values for `supercondition_factor` result in better agreement with the text but a narrower variety of generated images. Every image token is sampled from the `top_k` most probable tokens. The largest logit is subtracted from the logits to avoid infs. The logits are then divided by the `temperature`. If `is_seamless` is true, the images grid will be tiled in token space not pixel space.
If the model is being used interactively (e.g. in a notebook) `generate_image_stream` can be used to generate a stream of images as the model is decoding. The detokenizer adds a slight delay for each image. Set `progressive_outputs` to `True` to enable this. An example is implemented in the colab.