diffsynth-engine 0.3.0__tar.gz → 0.3.2__tar.gz

{diffsynth_engine-0.3.0 → diffsynth_engine-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.3.0
+Version: 0.3.2
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent

{diffsynth_engine-0.3.0 → diffsynth_engine-0.3.2}/diffsynth_engine/pipelines/base.py

@@ -151,11 +151,11 @@ class BasePipeline:
         noise = torch.randn(shape, generator=generator, device=device, dtype=dtype)
         return noise
 
-    def encode_image(self, image: torch.Tensor) -> torch.Tensor:
+    def encode_image(
+        self, image: torch.Tensor, tiled: bool = False, tile_size: int = 64, tile_stride: int = 32
+    ) -> torch.Tensor:
         image = image.to(self.device, self.vae_encoder.dtype)
-        latents = self.vae_encoder(
-            image, tiled=self.vae_tiled, tile_size=self.vae_tile_size, tile_stride=self.vae_tile_stride
-        )
+        latents = self.vae_encoder(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
         return latents
 
     def decode_image(self, latent: torch.Tensor) -> torch.Tensor:
@@ -187,7 +187,7 @@ class BasePipeline:
             self.load_models_to_device(["vae_encoder"])
             noise = latents
             image = self.preprocess_image(input_image).to(device=self.device, dtype=self.dtype)
-            latents = self.encode_image(image, tiled, tile_size, tile_stride)
+            latents = self.encode_image(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
             init_latents = latents.clone()
             latents = self.sampler.add_noise(latents, noise, sigma_start)
         else:
@@ -196,7 +196,8 @@ class BasePipeline:
             # if you have any questions about this, please ask @dizhipeng.dzp for more details
             latents = latents * sigmas[0] / ((sigmas[0] ** 2 + 1) ** 0.5)
             init_latents = latents.clone()
-        sigmas, timesteps = sigmas.to(device=self.device), timesteps.to(self.device)
+        sigmas, timesteps = sigmas.to(device=self.device, dtype=self.dtype), timesteps.to(device=self.device, dtype=self.dtype)
+        init_latents, latents = init_latents.to(device=self.device, dtype=self.dtype), latents.to(device=self.device, dtype=self.dtype)
         return init_latents, latents, sigmas, timesteps
 
     def eval(self):

{diffsynth_engine-0.3.0 → diffsynth_engine-0.3.2}/diffsynth_engine/pipelines/flux_image.py

@@ -676,7 +676,8 @@ class FluxImagePipeline(BasePipeline):
                 num_inference_steps, mu=mu, sigma_min=1 / num_inference_steps, sigma_max=1.0
             )
             init_latents = latents.clone()
-        sigmas, timesteps = sigmas.to(device=self.device), timesteps.to(self.device)
+        sigmas, timesteps = sigmas.to(device=self.device, dtype=self.dtype), timesteps.to(device=self.device, dtype=self.dtype)
+        init_latents, latents = init_latents.to(device=self.device, dtype=self.dtype), latents.to(device=self.device, dtype=self.dtype)
         return init_latents, latents, sigmas, timesteps
 
     def prepare_masked_latent(self, image: Image.Image, mask: Image.Image | None, height: int, width: int):

{diffsynth_engine-0.3.0 → diffsynth_engine-0.3.2}/diffsynth_engine.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: diffsynth_engine
-Version: 0.3.0
+Version: 0.3.2
 Author: MuseAI x ModelScope
 Classifier: Programming Language :: Python :: 3
 Classifier: Operating System :: OS Independent