Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -62,13 +62,13 @@ class DiffusionSampler:
|
|
62 |
self.alphas_cumprod_prev = torch.cat([torch.tensor([1.0]), self.alphas_cumprod[:-1]])
|
63 |
self.posterior_variance = self.betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
|
64 |
|
65 |
-
# Move to device
|
66 |
-
self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(self.device)
|
67 |
-
self.sqrt_one_minus_alpha_cumprod = self.sqrt_one_minus_alpha_cumprod.to(self.device)
|
68 |
-
self.sqrt_recip_alphas = self.sqrt_recip_alphas.to(self.device)
|
69 |
-
self.betas = self.betas.to(self.device)
|
70 |
-
self.posterior_variance = self.posterior_variance.to(self.device)
|
71 |
-
|
72 |
def load_vae(self):
|
73 |
"""Load VAE model (done lazily to save memory until needed)"""
|
74 |
if self.vae is None:
|
@@ -102,6 +102,13 @@ class DiffusionSampler:
|
|
102 |
|
103 |
# Start with random latents
|
104 |
latents = torch.randn((num_samples, 4, 32, 32), device=self.device, dtype=torch.float16)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
|
106 |
# Use classifier-free guidance for better quality
|
107 |
cfg_scale = 2.5
|
@@ -118,10 +125,10 @@ class DiffusionSampler:
|
|
118 |
|
119 |
t = torch.full((num_samples,), t_val, device=self.device, dtype=torch.long)
|
120 |
|
121 |
-
sqrt_recip_alphas_t =
|
122 |
-
sqrt_one_minus_alphas_cumprod_t =
|
123 |
-
beta_t =
|
124 |
-
posterior_variance_t =
|
125 |
|
126 |
# Get noise prediction with classifier-free guidance
|
127 |
eps_theta_cond = model(latents, t, g_cond, s_cond)
|
@@ -130,7 +137,7 @@ class DiffusionSampler:
|
|
130 |
|
131 |
# Update latents
|
132 |
mean = sqrt_recip_alphas_t * (latents - (beta_t / sqrt_one_minus_alphas_cumprod_t) * eps_theta)
|
133 |
-
noise = torch.randn_like(latents)
|
134 |
if t_val == 0:
|
135 |
latents = mean
|
136 |
else:
|
|
|
62 |
self.alphas_cumprod_prev = torch.cat([torch.tensor([1.0]), self.alphas_cumprod[:-1]])
|
63 |
self.posterior_variance = self.betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
|
64 |
|
65 |
+
# Move to device
|
66 |
+
self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(self.device)
|
67 |
+
self.sqrt_one_minus_alpha_cumprod = self.sqrt_one_minus_alpha_cumprod.to(self.device)
|
68 |
+
self.sqrt_recip_alphas = self.sqrt_recip_alphas.to(self.device)
|
69 |
+
self.betas = self.betas.to(self.device)
|
70 |
+
self.posterior_variance = self.posterior_variance.to(self.device)
|
71 |
+
|
72 |
def load_vae(self):
|
73 |
"""Load VAE model (done lazily to save memory until needed)"""
|
74 |
if self.vae is None:
|
|
|
102 |
|
103 |
# Start with random latents
|
104 |
latents = torch.randn((num_samples, 4, 32, 32), device=self.device, dtype=torch.float16)
|
105 |
+
|
106 |
+
# Convert diffusion parameters to half precision for compatibility
|
107 |
+
sqrt_alphas_cumprod_half = self.sqrt_alphas_cumprod.half()
|
108 |
+
sqrt_one_minus_alpha_cumprod_half = self.sqrt_one_minus_alpha_cumprod.half()
|
109 |
+
sqrt_recip_alphas_half = self.sqrt_recip_alphas.half()
|
110 |
+
betas_half = self.betas.half()
|
111 |
+
posterior_variance_half = self.posterior_variance.half()
|
112 |
|
113 |
# Use classifier-free guidance for better quality
|
114 |
cfg_scale = 2.5
|
|
|
125 |
|
126 |
t = torch.full((num_samples,), t_val, device=self.device, dtype=torch.long)
|
127 |
|
128 |
+
sqrt_recip_alphas_t = sqrt_recip_alphas_half[t].view(-1, 1, 1, 1)
|
129 |
+
sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alpha_cumprod_half[t].view(-1, 1, 1, 1)
|
130 |
+
beta_t = betas_half[t].view(-1, 1, 1, 1)
|
131 |
+
posterior_variance_t = posterior_variance_half[t].view(-1, 1, 1, 1)
|
132 |
|
133 |
# Get noise prediction with classifier-free guidance
|
134 |
eps_theta_cond = model(latents, t, g_cond, s_cond)
|
|
|
137 |
|
138 |
# Update latents
|
139 |
mean = sqrt_recip_alphas_t * (latents - (beta_t / sqrt_one_minus_alphas_cumprod_t) * eps_theta)
|
140 |
+
noise = torch.randn_like(latents,dtype=torch.float16)
|
141 |
if t_val == 0:
|
142 |
latents = mean
|
143 |
else:
|