test

Author: yiyixuxu

src/diffusers/pipelines/mochi/pipeline_mochi.py

@@ -660,11 +660,10 @@ def __call__(
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                 timestep = t.expand(latents.shape[0]).to(latents.dtype)
 
-                print(t)
                 noise_pred = self.transformer(
                     hidden_states=latent_model_input,
                     encoder_hidden_states=prompt_embeds,
-                    timestep=1000 - timestep,
+                    timestep=timestep,
                     encoder_attention_mask=prompt_attention_mask,
                     return_dict=False,
                 )[0]

src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

@@ -205,9 +205,15 @@ def set_timesteps(
         sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
         timesteps = sigmas * self.config.num_train_timesteps
 
-        self.timesteps = timesteps.to(device=device)
-        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+        if self.config.invert_sigmas:
+            sigmas = 1.0 - sigmas
+            timesteps = sigmas * self.config.num_train_timesteps
+            sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)])
+        else:
+            sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
 
+        self.timesteps = timesteps.to(device=device)
+        self.sigmas = sigmas
         self._step_index = None
         self._begin_index = None
 
@@ -295,10 +301,6 @@ def step(
         sigma = self.sigmas[self.step_index]
         sigma_next = self.sigmas[self.step_index + 1]
 
-        if self.config.invert_sigmas:
-            print("inverting")
-            sigma, sigma_next = sigma_next, sigma
-
         prev_sample = sample + (sigma_next - sigma) * model_output
 
         # Cast sample back to model compatible dtype