CompVis

/

stable-diffusion-v-1-4-original

like 2.66k

From what I understand,
'sd-v1-4_ckpt' contains just the weights and biases for inference.
'sd-v1-4-full-ema_ckpt' contains the weights and biases + momentum and other training related metadata, so you could continue the training steps.
Might be related to this: https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage

so regular users should just use sd-v1-4_ckpt and not full-ema, right?

so regular users should just use sd-v1-4_ckpt and not full-ema, right?

You could use both for inference, it's just that 'sd-v1-4_ckpt' is smaller in size so faster to download, and save some disk space.
For inference, both should be almost the same, but 'sd-v1-4-full-ema_ckpt' should produce a better result according to these:
https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage

For inference, both should be almost the same, but 'sd-v1-4-full-ema_ckpt' should produce a better result according to these:
https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage

But its about training right?

so regular users should just use sd-v1-4_ckpt and not full-ema, right?

You could use both for inference, it's just that 'sd-v1-4_ckpt' is smaller in size so faster to download, and save some disk space.
For inference, both should be almost the same, but 'sd-v1-4-full-ema_ckpt' should produce a better result according to these:
https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage

has it been tested on same seed/same precision and is full-ema producing better results?

@irkson0101 i've tested with a bunch of prompts.
imo sometime full-ema has a better output, sometime the other one is the winner.

really i'm not sure if one is actually better.

so regular users should just use sd-v1-4_ckpt and not full-ema, right?

You could use both for inference, it's just that 'sd-v1-4_ckpt' is smaller in size so faster to download, and save some disk space.
For inference, both should be almost the same, but 'sd-v1-4-full-ema_ckpt' should produce a better result according to these:
https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage

has it been tested on same seed/same precision and is full-ema producing better results?

Just stumbled across this interesting comparison showing both side-by-side, the difference seems negligible.

after I download the weights file. I copy it the stable-diffusion-v1 folder and rename it model.ckpt
then I close and press the validate button I get an error "NO module named 'antlr4' can someone please let me know what i have don't wrong? all other steps said they were successful