DreamBooth training in under 8 GB VRAM and textual inversion under 6 GB!

[ZeroCool22]

https://www.reddit.com/r/StableDiffusion/comments/xwdj79/dreambooth_training_in_under_8_gb_vram_and/

https://github.com/Ttl/diffusers/tree/dreambooth_deepspeed

Ttl/diffusers@9ea0078

The drawback is of course that now the training requires significantly more RAM (about 25 GB).

[Ehplodor]

This is crazy. Only have 16 GB CPU ram available, though... 😓

[0xdevalias]

Potentially Related:

• Implement new paper: Dreambooth-StableDiffusion, Google Imagen based Textual Inversion alternative #914
• Running AUTOMATIC1111 / stable-diffusion-webui with Dreambooth fine-tuned models #1429
• [Feature request] Dreambooth deepspeed #1734
• [Feature Request]: Dreambooth on 8GB VRam GPU (holy grail) #3586
• Dreambooth #2002
  • Dreambooth #2002 (comment)

    • Closing, opening new PR to squash commits and make it clean.

• Dreambooth: Ready to go! #3995
  • Dreambooth: Ready to go! #3995 (comment)

    • Please give this a look and merge.