Version 3.6 improves ease of use and has heavy shift towards more illustrative qualities. Level of photorealism still exists within the model but if you are looking for photoreal model, probably better stay with the version 1.
Knowledge of anatomy greatly improved but still lacking certain details.
Proletarian is single GPU finetune of Stable Cascade 3B to basically find out viability of tuning Wuerstchen architecture as solo contributor. It's trained with OneTrainer mostly using Adafactor, because it allows for fairly large batches even for single GPU.
For generating, please use the separately shared workflow or your own, place UNET into UNET folder of your ComfyUI installation and Text Encoder in the CLIP folder of your ComfyUI installation. Use Stage B and Stage A from Stability AI Huggingface account.
It is pretty flexible with subject matter and styles and has been trained on a dataset of around 40k high-quality images. Prompt following is not on the level of premium SDXL finetunes, but considerably better than any of SD1.5 tunes or merges I have touched. Stable Cascade listens well to natural language, but as I have trained with Danbooru tags, the model does pretty well with tagging style prompt as well.
There are some baked in quality tags like "apg_v3", "plt_2", "pda_v10" and "pdxl_v6" that are trained on synthetic data, please feel free to experiment with them. There are also bunch of artist names added to Stable Cascades already a large catalogue, but I won't be releasing them as they are not at all representative of the artists talent.
The model does some basic nudity but is clearly not fully confident with adult scenes.
I will keep training the model further and collecting a bigger dataset, but if you want to support the progress feel free to support me on Patreon.
Description
Proletarian is single GPU finetune of Stable Cascade to basically find out viability of tuning Wuerstchen architecture as solo contributor. It's trained with OneTrainer mostly using Adafactor, because it allows for fairly large batches even for single GPU.
It is pretty flexible with subject matter and styles and has been trained on a dataset of around 40k high-quality images. Prompt following is not on the level of premium SDXL finetunes, but considerably better than any of SD1.5 tunes or merges I have touched. Stable Cascade listens well to natural language, but as I have trained with Danbooru tags, the model does pretty well with tagging style prompt as well.
Resolutions are pretty flexible, if the output resolution is divisible by the compression factor it should be fine, duplication of features starts from 1536 upward if compression is not adjusted. Steps for samples are with AYS at 30, but that is likely overkill and model works fine with less.
There are some baked in quality tags like "apg_v3", "plt_2", "pda_v10" and "pdxl_v6" that are trained on synthetic data, please feel free to experiment with them. There are also bunch of artist names added to Stable Cascades already a large catalogue, but I won't be releasing them as they are not at all representative of the artists talent.
The model does some basic nudity but is clearly not fully confident with adult scenes.
I will keep training the model further and collecting a bigger dataset, but if you want to support the progress feel free to support me on Patreon.
FAQ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.