kuprel/min-dalle

Stripped down to core inference with minimal dependencies (numpy, requests, pillow, torch), it uses a two-stage architecture combining a text encoder with an autoregressive image token decoder converted from the original Flax implementation. Supports progressive streaming outputs, configurable precision modes (float32/float16/bfloat16), and sampling controls like temperature and top-k filtering for balancing image quality against diversity.

3,499 stars and 813 monthly downloads. No commits in the last 6 months. Available on PyPI.