search

ExponentialML / AnimateDiff-MotionDirector

MotionDirector Training For AnimateDiff. Train a MotionLoRA and run it on any compatible AnimateDiff UI.
Apache License 2.0
291 stars 28 forks source link

Is it necessary to download the whole SD 1.5 repo ? #24

Open tonynoce opened 8 months ago

tonynoce commented 8 months ago

Been trying to make this work since yesterday ( still havent given up yet ), but yesterday I ran out of disk space, the models/StableDiffusion got into the 80gb.

Is it possible to ran the train.py with a more pruned version of the repo ?

kuzman123 commented 7 months ago

Same Q here.. I've stopped download od 40 gigs..

user-vm commented 6 months ago

I don't think you need any of the safetensors files, and you don't need the fp16 or ema_only variants of the .bin files. I think just these files should work (some might be extra):

./model_index.json
./scheduler/scheduler_config.json
./vae/diffusion_pytorch_model.bin
./vae/config.json
./safety_checker/config.json
./unet/diffusion_pytorch_model.bin
./unet/config.json
./text_encoder/pytorch_model.bin
./text_encoder/config.json
./feature_extractor/preprocessor_config.json
./tokenizer/special_tokens_map.json
./tokenizer/tokenizer_config.json
./tokenizer/vocab.json
./tokenizer/merges.txt

Also, if you've used SD1.5 with diffusers before, you may already have the required files on your system somewhere (like in ~/.cache/huggingface/hub on Linux).