Sigma Proletarian is small finetune of Pixart Sigma for 5 epochs with dataset of roughly 50k images. It is likely still severely undertrained, but strictly different from the base model. To use the model you need T5 text encoder of your choice. Works with SD3 text encoder if set up for that, can be run with lowvram mode with 6Gb VRAM in ComfyUI. There's some drop in quality when using small quant version but it is still usable. Model follows natural language prompting well, but can also use danbooru style tags not too badly. There are few quality tags (apg_v3, pda_v10, pdxl_v6, hw_5) that are trained with small set of synthetic data to enforce quality by putting it in front of the prompt. There's roughly 50 artist names in the training dataset, but their style prompted is not descriptive of the style.

Files with SHA256: 8986f7b241c6982e0e008625c1dfb31e3c9fd69b8bad19dfdb636cedda4cdb3b