nodes/modelscope-scepter
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

update v0.0.3

Browse Source
This commit is contained in:
hanzhn
2024-02-07 14:57:28 +08:00
parent 9999e0e1f9
commit 3440ec7c38
9 changed files with 439 additions and 299 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scepter/modules/model/network/diffusion/diffusion.py
+12 -9
View File
@@ -54,7 +54,8 @@ class GaussianDiffusion(object):
guide_rescale=None,
clamp=None,
percentile=None,
cat_uc=False):
cat_uc=False,
**kwargs):
"""
Apply one step of denoising from the posterior distribution q(x_s | x_t, x0).
Since x0 is not available, estimate the denoising results using the learned
@@ -79,7 +80,7 @@ class GaussianDiffusion(object):
# prediction
if guide_scale is None:
assert isinstance(model_kwargs, dict)
out = model(xt, t=t, **model_kwargs)
out = model(xt, t=t, **model_kwargs, **kwargs)
else:
# classifier-free guidance (arXiv:2207.12598)
# model_kwargs[0]: conditional kwargs
@@ -87,7 +88,7 @@ class GaussianDiffusion(object):
assert isinstance(model_kwargs, list) and len(model_kwargs) == 2
if guide_scale == 1.:
out = model(xt, t=t, **model_kwargs[0])
out = model(xt, t=t, **model_kwargs[0], **kwargs)
else:
if cat_uc:
@@ -111,11 +112,12 @@ class GaussianDiffusion(object):
all_model_kwargs[key], value)
all_out = model(xt.repeat(2, 1, 1, 1),
t=t.repeat(2),
**all_model_kwargs)
**all_model_kwargs,
**kwargs)
y_out, u_out = all_out.chunk(2)
else:
y_out = model(xt, t=t, **model_kwargs[0])
u_out = model(xt, t=t, **model_kwargs[1])
y_out = model(xt, t=t, **model_kwargs[0], **kwargs)
u_out = model(xt, t=t, **model_kwargs[1], **kwargs)
out = u_out + guide_scale * (y_out - u_out)
# rescale the output according to arXiv:2305.08891
@@ -262,7 +264,8 @@ class GaussianDiffusion(object):
guide_rescale,
clamp,
percentile,
cat_uc=cat_uc)[-2]
cat_uc=cat_uc,
**kwargs)[-2]
# collect intermediate outputs
if return_intermediate == 'xt':
@@ -467,8 +470,8 @@ class GaussianDiffusion(object):
t = self._sigma_to_t(sigma).repeat(len(xt)).round().long()
x0 = self.denoise(xt * c_in, t, None, model, model_kwargs,
guide_scale, guide_rescale, clamp,
percentile)[-2]
guide_scale, guide_rescale, clamp, percentile,
**kwargs)[-2]
# collect intermediate outputs
if return_intermediate == 'xt':
intermediates.append(xt)