Encode batches separately

Significantly reduces VRAM.
This makes encoding more inline with how decoding currently functions.
This commit is contained in:
catboxanon 2023-08-13 04:16:48 -04:00
parent da80d649fd
commit 822597db49
1 changed files with 9 additions and 1 deletions

View File

@ -92,7 +92,15 @@ def images_tensor_to_samples(image, approximation=None, model=None):
model = shared.sd_model model = shared.sd_model
image = image.to(shared.device, dtype=devices.dtype_vae) image = image.to(shared.device, dtype=devices.dtype_vae)
image = image * 2 - 1 image = image * 2 - 1
x_latent = model.get_first_stage_encoding(model.encode_first_stage(image)) if len(image) > 1:
x_latent = torch.stack([
model.get_first_stage_encoding(
model.encode_first_stage(torch.unsqueeze(img, 0))
)[0]
for img in image
])
else:
x_latent = model.get_first_stage_encoding(model.encode_first_stage(image))
return x_latent return x_latent