test5_4_2800 #safetensors
new version is available
generate photorealistic cp with this finetune
this finetune has been trained with ~37 illegal pictures
test5_4_1700.safetensors
#checkpoint #safetensors #finetuning #stablediffusion #sd #sd15
test5_4_900.safetensors
quite some improvement over previous version
ill do some gens and post samples soon
after disabling [] cache latents, im gettig pretty good ~2.3s/it. although its still showing vram usage about same as before
#sd15 #finetuning
this is a finetune of basilmix with 20 images
still highly experimental
can occasionally result to decent gens
here are some cherry picked results
test5_3_1600.safetensors
upgraded ram from 16gb to 32gb
now i can allocate 20gb for wsl
way more stabile training speed and overall computer usability when swap is not needed
Some comments on training with 8gb RAM in wsl:
- xformers is needed, but installing it May break everything. Make sure bitsandbytes works with your cuda version. Note that .sh script overrides installed packages. Start with:
python webui.py --xformers
Default settings in dreambooth works ok. Make sure xformers is selected.
Resuming from previously finished checkpoint wants massive amounts of RAM. Starting training from og SD 1.5 works best. So far no success trying to train from Basil mix checkpoint.
I have set up wsl with 9GB RAM, 6gb swap
#sd15 #stablediffusion #training #sdwebui
did fist finetuning from og sd1.5 model. not very special results yet, but its possible to finetune sd1.5 with just 8gb vram. it took maybe an hour using 21 pictures.
