DPMSolverMultistep add rescale_betas_zero_snr

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -70,6 +70,42 @@ def alpha_bar_fn(t):
         betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+    Args:
+        betas (`torch.FloatTensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.FloatTensor`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = torch.cat([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
 
 class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     """
@@ -144,6 +180,10 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             An offset added to the inference steps. You can use a combination of `offset=1` and
             `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
             Diffusion.
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -173,6 +213,7 @@ def __init__(
         variance_type: Optional[str] = None,
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
+        rescale_betas_zero_snr: bool = False,
     ):
         if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
             deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
@@ -191,8 +232,17 @@ def __init__(
         else:
             raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
 
+        if rescale_betas_zero_snr:
+            self.betas = rescale_zero_terminal_snr(self.betas)
+
         self.alphas = 1.0 - self.betas
         self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        if rescale_betas_zero_snr:
+            # Close to 0 without being 0 so first sigma is not inf
+            # FP16 smallest positive subnormal works well here
+            self.alphas_cumprod[-1] = 2**-24
+
         # Currently we only support VP-type noise schedule
         self.alpha_t = torch.sqrt(self.alphas_cumprod)
         self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
@@ -895,9 +945,12 @@ def step(
             self.model_outputs[i] = self.model_outputs[i + 1]
         self.model_outputs[-1] = model_output
 
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+
         if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
             noise = randn_tensor(
-                model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+                model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32
             )
         else:
             noise = None
@@ -912,6 +965,9 @@ def step(
         if self.lower_order_nums < self.config.solver_order:
             self.lower_order_nums += 1
 
+        # Cast sample back to expected dtype
+        prev_sample = prev_sample.to(model_output.dtype)
+
         # upon completion increase step index by one
         self._step_index += 1
 

tests/schedulers/test_scheduler_dpm_multi.py

@@ -213,6 +213,10 @@ def test_inference_steps(self):
         for num_inference_steps in [1, 2, 3, 5, 10, 50, 100, 999, 1000]:
             self.check_over_forward(num_inference_steps=num_inference_steps, time_step=0)
 
+    def test_rescale_betas_zero_snr(self):
+        for rescale_betas_zero_snr in [True, False]:
+            self.check_over_configs(rescale_betas_zero_snr=rescale_betas_zero_snr)
+
     def test_full_loop_no_noise(self):
         sample = self.full_loop()
         result_mean = torch.mean(torch.abs(sample))