improve: bump basicsr version

perceptor/models/stable_diffusion/predictions.py

@@ -44,9 +44,6 @@ def from_sigmas(self):
 
     @property
     def denoised_latents(self):
-        # return (
-        #     latents - self.sqrt_one_minus_alphas_cumprod(index) * eps
-        # ) / self.alphas_cumprod(index).sqrt()
         return (
             self.from_diffused_latents - self.from_sigmas * self.predicted_noise
         ) / self.from_alphas.clamp(min=1e-7)

perceptor/models/stable_diffusion/stable_diffusion.py

@@ -1,5 +1,6 @@
 from typing import Optional
 from contextlib import contextmanager
+import copy
 import torch
 import lantern
 from transformers import CLIPTokenizer, CLIPTextModel, logging
@@ -133,7 +134,7 @@ def finetuneable_vae(self):
         with diffusion_model.finetuneable_vae():
             images = diffusion_model.decode(latents)
         """
-        state_dict = self.vae.state_dict()
+        state_dict = copy.deepcopy(self.vae.state_dict())
         try:
             for parameter, requires_grad in zip(
                 self.vae.parameters(), self.vae_original_requires_grads

perceptor/models/velocity_diffusion/model_urls.py

@@ -1,7 +1,7 @@
 model_urls = {
-    "yfcc_2": "https://the-eye.eu/public/AI/models/v-diffusion/yfcc_2.pth",
-    "yfcc_1": "https://the-eye.eu/public/AI/models/v-diffusion/yfcc_1.pth",
-    "cc12m_1": "https://the-eye.eu/public/AI/models/v-diffusion/cc12m_1.pth",
-    "cc12m_1_cfg": "https://the-eye.eu/public/AI/models/v-diffusion/cc12m_1_cfg.pth",
+    "yfcc_2": "https://s3.eu-central-1.wasabisys.com/nextml-model-data/velocity-diffusion/yfcc_2.pth",
+    "yfcc_1": "https://s3.eu-central-1.wasabisys.com/nextml-model-data/velocity-diffusion/yfcc_1.pth",
+    "cc12m_1": "https://s3.eu-central-1.wasabisys.com/nextml-model-data/velocity-diffusion/cc12m_1.pth",
+    "cc12m_1_cfg": "https://s3.eu-central-1.wasabisys.com/nextml-model-data/velocity-diffusion/cc12m_1_cfg.pth",
     "wikiart": "https://the-eye.eu/public/AI/models/v-diffusion/wikiart_256.pth",
 }

perceptor/models/velocity_diffusion/predictions.py

@@ -104,21 +104,6 @@ def step(self, to_ts, eta=0.0):
         return diffusion_space.decode(to_diffused_xs)
 
     def correction(self, previous: "Predictions"):
-        # k-diffusion has alphas=1 always so this should not work
-        # corrected_diffused_xs = (
-        #     previous.from_diffused_xs
-        #     + (self.from_sigmas - previous.from_sigmas) * (self.eps - previous.eps) / 2
-        # )
-        # return diffusion_space.decode(corrected_diffused_xs)
-
-        # blurry
-        # return Predictions(
-        #     from_diffused_images=previous.from_diffused_images,
-        #     from_ts=previous.from_ts,
-        #     velocities=(self.velocities + previous.velocities) / 2,
-        # )
-
-        # looks ok but no apparent difference from double budget
         return previous.forced_denoised(
             (self.denoised_images + previous.denoised_images) / 2
         )
@@ -127,29 +112,9 @@ def reverse_step(self, to_ts):
         if (torch.as_tensor(self.from_ts) > torch.as_tensor(to_ts)).any():
             raise ValueError("from_ts must be less than to_ts")
 
-        # replaced_noise_sigma = (
-        #     self.sigmas(to_t) ** 2 - self.sigmas(from_t) ** 2
-        # ).sqrt()
-        # to_eps = (
-        #     from_eps * self.sigmas(from_t)
-        #     + torch.randn_like(from_eps) * replaced_noise_sigma
-        # ) / self.sigmas(to_t)
-        # to_eps = eps
-        # to_diffused = self.diffuse(from_denoised, to_t, noise=to_eps)
-        # return to_diffused
-
         to_alphas, to_sigmas = self.alphas(to_ts), self.sigmas(to_ts)
         return self.denoised_xs * to_alphas + self.predicted_noise * to_sigmas
 
-    def resample(self, resample_ts):
-        """
-        Harmonizing resampling from https://github.com/andreas128/RePaint
-        """
-        return diffusion_space.decode(
-            self.denoised_xs * self.from_alphas
-            + self.resample_noise(resample_ts) * self.from_sigmas
-        )
-
     def resample_noise(self, resample_ts):
         if (torch.as_tensor(self.from_ts) < torch.as_tensor(resample_ts)).any():
             raise ValueError("from_ts must be greater than resample_ts")
@@ -161,6 +126,15 @@ def resample_noise(self, resample_ts):
         )  # fmt: skip
         return resampled_noise_sigma / self.from_sigmas
 
+    def resample(self, resample_ts):
+        """
+        Harmonizing resampling from https://github.com/andreas128/RePaint
+        """
+        return diffusion_space.decode(
+            self.denoised_xs * self.from_alphas
+            + self.resample_noise(resample_ts) * self.from_sigmas
+        )
+
     def noisy_reverse_step(self, to_ts):
         to_alphas, to_sigmas = self.alphas(to_ts), self.sigmas(to_ts)