鈿狅笍 This is not an SDXL or SD 1.5 or Pony model
This is a fine tune of PixArt Sigma 1024 on the training dataset I use for training PixelWave.
Version 01 was trained on 22k mostly non AI images for 50 epochs, and then merged with the 00 checkpoint to help improve the overall output of the model.
The 00 version is only 5 epochs. I included over 70k images generated with SDXL version of PixelWave in the training.
I'm still looking for the best settings. More steps usually gives a better image.
You need the extra models node package for comfy. https://github.com/city96/ComfyUI_ExtraModels
If you have the vram you can set the T5 loader to use gpu instead of cpu ram.
T5 files go in the models/t5 folder: https://huggingface.co/PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers/tree/main/text_encoder
Description
FAQ
Comments (6)
Looks good.
i'm finding this to be the best sigma finetune so far. The fact that you did not train much turned out great
can run on a1111?
nothing can run on auto1111 for a long time. You can run it in sd-next, comfyui and stable swarm ui.
This checkpoint is tiny! I'm used to ~6.5gb SDXL models. Do you think this checkpoint and the t5 encoder would fit on an 8gb GPU?
The T5 encoder can run on your CPU. I can run Pixart Sigma on my laptop's 6GB 3060 fine.