Add stochastic sampling to DPM solver (SDE)

src/refiners/foundationals/latent_diffusion/solvers/ddim.py

@@ -41,6 +41,8 @@ class DDIM(Solver):
         """
         if params and params.model_prediction_type not in (ModelPredictionType.NOISE, None):
             raise NotImplementedError
+        if params and params.sde_variance != 0.0:
+            raise NotImplementedError("DDIM does not support sde_variance != 0.0 yet")
 
         super().__init__(
             num_inference_steps=num_inference_steps,

src/refiners/foundationals/latent_diffusion/solvers/dpm.py

@@ -2,7 +2,8 @@ import dataclasses
 from collections import deque
 
 import numpy as np
-from torch import Generator, Tensor, device as Device, dtype as Dtype, exp, float32, tensor
+import torch
+from torch import Generator, Tensor, device as Device, dtype as Dtype, float32, tensor
 
 from refiners.foundationals.latent_diffusion.solvers.solver import (
     BaseSolverParams,
@@ -51,6 +52,8 @@ class DPMSolver(Solver):
         """
         if params and params.model_prediction_type not in (ModelPredictionType.NOISE, None):
             raise NotImplementedError
+        if params and params.sde_variance not in (0.0, 1.0):
+            raise NotImplementedError("DPMSolver only supports sde_variance=0.0 or 1.0")
 
         super().__init__(
             num_inference_steps=num_inference_steps,
@@ -93,7 +96,9 @@ class DPMSolver(Solver):
         np_space = np.linspace(offset, max_timestep, self.num_inference_steps + 1).round().astype(int)[1:]
         return tensor(np_space).flip(0)
 
-    def dpm_solver_first_order_update(self, x: Tensor, noise: Tensor, step: int) -> Tensor:
+    def dpm_solver_first_order_update(
+        self, x: Tensor, noise: Tensor, step: int, sde_noise: Tensor | None = None
+    ) -> Tensor:
         """Applies a first-order backward Euler update to the input data `x`.
 
         Args:
@@ -115,11 +120,21 @@ class DPMSolver(Solver):
         previous_noise_std = self.noise_std[previous_timestep]
         current_noise_std = self.noise_std[current_timestep]
 
-        factor = exp(-(previous_ratio - current_ratio)) - 1.0
+        ratio_delta = current_ratio - previous_ratio
-        denoised_x = (previous_noise_std / current_noise_std) * x - (factor * previous_scale_factor) * noise
-        return denoised_x
 
-    def multistep_dpm_solver_second_order_update(self, x: Tensor, step: int) -> Tensor:
+        if sde_noise is None:
+            return (previous_noise_std / current_noise_std) * x + (
+                1.0 - torch.exp(ratio_delta)
+            ) * previous_scale_factor * noise
+
+        factor = 1.0 - torch.exp(2.0 * ratio_delta)
+        return (
+            (previous_noise_std / current_noise_std) * torch.exp(ratio_delta) * x
+            + previous_scale_factor * factor * noise
+            + previous_noise_std * torch.sqrt(factor) * sde_noise
+        )
+
+    def multistep_dpm_solver_second_order_update(self, x: Tensor, step: int, sde_noise: Tensor | None = None) -> Tensor:
         """Applies a second-order backward Euler update to the input data `x`.
 
         Args:
@@ -147,13 +162,23 @@ class DPMSolver(Solver):
         estimation_delta = (current_data_estimation - next_data_estimation) / (
             (current_ratio - next_ratio) / (previous_ratio - current_ratio)
         )
-        factor = exp(-(previous_ratio - current_ratio)) - 1.0
+        ratio_delta = current_ratio - previous_ratio
-        denoised_x = (
+
-            (previous_noise_std / current_noise_std) * x
+        if sde_noise is None:
-            - (factor * previous_scale_factor) * current_data_estimation
+            factor = 1.0 - torch.exp(ratio_delta)
-            - 0.5 * (factor * previous_scale_factor) * estimation_delta
+            return (
+                (previous_noise_std / current_noise_std) * x
+                + previous_scale_factor * factor * current_data_estimation
+                + 0.5 * previous_scale_factor * factor * estimation_delta
+            )
+
+        factor = 1.0 - torch.exp(2.0 * ratio_delta)
+        return (
+            (previous_noise_std / current_noise_std) * torch.exp(ratio_delta) * x
+            + previous_scale_factor * factor * current_data_estimation
+            + 0.5 * previous_scale_factor * factor * estimation_delta
+            + previous_noise_std * torch.sqrt(factor) * sde_noise
         )
-        return denoised_x
 
     def __call__(self, x: Tensor, predicted_noise: Tensor, step: int, generator: Generator | None = None) -> Tensor:
         """Apply one step of the backward diffusion process.
@@ -175,11 +200,20 @@ class DPMSolver(Solver):
         assert self.first_inference_step <= step < self.num_inference_steps, "invalid step {step}"
 
         current_timestep = self.timesteps[step]
-        scale_factor, noise_ratio = self.cumulative_scale_factors[current_timestep], self.noise_std[current_timestep]
+        scale_factor = self.cumulative_scale_factors[current_timestep]
+        noise_ratio = self.noise_std[current_timestep]
         estimated_denoised_data = (x - noise_ratio * predicted_noise) / scale_factor
         self.estimated_data.append(estimated_denoised_data)
+        variance = self.params.sde_variance
+        sde_noise = (
+            torch.randn(x.shape, generator=generator, device=x.device, dtype=x.dtype) * variance
+            if variance > 0.0
+            else None
+        )
 
         if step == self.first_inference_step or (self.last_step_first_order and step == self.num_inference_steps - 1):
-            return self.dpm_solver_first_order_update(x=x, noise=estimated_denoised_data, step=step)
+            return self.dpm_solver_first_order_update(
+                x=x, noise=estimated_denoised_data, step=step, sde_noise=sde_noise
+            )
 
-        return self.multistep_dpm_solver_second_order_update(x=x, step=step)
+        return self.multistep_dpm_solver_second_order_update(x=x, step=step, sde_noise=sde_noise)

src/refiners/foundationals/latent_diffusion/solvers/euler.py

@@ -36,6 +36,8 @@ class Euler(Solver):
         """
         if params and params.noise_schedule not in (NoiseSchedule.QUADRATIC, None):
             raise NotImplementedError
+        if params and params.sde_variance != 0.0:
+            raise NotImplementedError("Euler does not support sde_variance != 0.0 yet")
 
         super().__init__(
             num_inference_steps=num_inference_steps,

src/refiners/foundationals/latent_diffusion/solvers/solver.py

@@ -79,6 +79,7 @@ class BaseSolverParams:
     final_diffusion_rate: float | None
     noise_schedule: NoiseSchedule | None
     model_prediction_type: ModelPredictionType | None
+    sde_variance: float
 
 
 @dataclasses.dataclass(kw_only=True, frozen=True)
@@ -102,6 +103,7 @@ class SolverParams(BaseSolverParams):
     final_diffusion_rate: float | None = None
     noise_schedule: NoiseSchedule | None = None
     model_prediction_type: ModelPredictionType | None = None
+    sde_variance: float = 0.0
 
 
 @dataclasses.dataclass(kw_only=True, frozen=True)
@@ -113,6 +115,7 @@ class ResolvedSolverParams(BaseSolverParams):
     final_diffusion_rate: float
     noise_schedule: NoiseSchedule
     model_prediction_type: ModelPredictionType
+    sde_variance: float
 
 
 class Solver(fl.Module, ABC):
@@ -123,6 +126,19 @@ class Solver(fl.Module, ABC):
 
     This process is described using several parameters such as initial and final diffusion rates,
     and is encapsulated into a `__call__` method that applies a step of the diffusion process.
+
+    Attributes:
+        params: The common parameters for solvers. See `SolverParams`.
+        num_inference_steps: The number of inference steps to perform.
+        first_inference_step: The step to start the inference process from.
+        scale_factors: The scale factors used to denoise the input. These are called "betas" in other implementations,
+            and `1 - scale_factors` is called "alphas".
+        cumulative_scale_factors: The cumulative scale factors used to denoise the input. These are called "alpha_t" in
+            other implementations.
+        noise_std: The standard deviation of the noise used to denoise the input. This is called "sigma_t" in other
+            implementations.
+        signal_to_noise_ratios: The signal-to-noise ratios used to denoise the input. This is called "lambda_t" in other
+            implementations.
     """
 
     timesteps: Tensor
@@ -136,6 +152,7 @@ class Solver(fl.Module, ABC):
         final_diffusion_rate=1.2e-2,
         noise_schedule=NoiseSchedule.QUADRATIC,
         model_prediction_type=ModelPredictionType.NOISE,
+        sde_variance=0.0,
     )
 
     def __init__(

tests/e2e/test_diffusion.py

@@ -88,6 +88,11 @@ def expected_image_std_random_init(ref_path: Path) -> Image.Image:
     return _img_open(ref_path / "expected_std_random_init.png").convert("RGB")
 
 
+@pytest.fixture
+def expected_image_std_sde_random_init(ref_path: Path) -> Image.Image:
+    return _img_open(ref_path / "expected_std_sde_random_init.png").convert("RGB")
+
+
 @pytest.fixture
 def expected_image_std_random_init_euler(ref_path: Path) -> Image.Image:
     return _img_open(ref_path / "expected_std_random_init_euler.png").convert("RGB")
@@ -560,6 +565,24 @@ def sd15_std(
     return sd15
 
 
+@pytest.fixture
+def sd15_std_sde(
+    text_encoder_weights: Path, lda_weights: Path, unet_weights_std: Path, test_device: torch.device
+) -> StableDiffusion_1:
+    if test_device.type == "cpu":
+        warn("not running on CPU, skipping")
+        pytest.skip()
+
+    sde_solver = DPMSolver(num_inference_steps=30, last_step_first_order=True, params=SolverParams(sde_variance=1.0))
+    sd15 = StableDiffusion_1(device=test_device, solver=sde_solver)
+
+    sd15.clip_text_encoder.load_from_safetensors(text_encoder_weights)
+    sd15.lda.load_from_safetensors(lda_weights)
+    sd15.unet.load_from_safetensors(unet_weights_std)
+
+    return sd15
+
+
 @pytest.fixture
 def sd15_std_float16(
     text_encoder_weights: Path, lda_weights: Path, unet_weights_std: Path, test_device: torch.device
@@ -831,6 +854,33 @@ def test_diffusion_std_random_init(
     ensure_similar_images(predicted_image, expected_image_std_random_init)
 
 
+@no_grad()
+def test_diffusion_std_sde_random_init(
+    sd15_std_sde: StableDiffusion_1, expected_image_std_sde_random_init: Image.Image, test_device: torch.device
+):
+    sd15 = sd15_std_sde
+
+    prompt = "a cute cat, detailed high-quality professional image"
+    negative_prompt = "lowres, bad anatomy, bad hands, cropped, worst quality"
+    clip_text_embedding = sd15.compute_clip_text_embedding(text=prompt, negative_text=negative_prompt)
+
+    sd15.set_inference_steps(50)
+
+    manual_seed(2)
+    x = torch.randn(1, 4, 64, 64, device=test_device)
+
+    for step in sd15.steps:
+        x = sd15(
+            x,
+            step=step,
+            clip_text_embedding=clip_text_embedding,
+            condition_scale=7.5,
+        )
+    predicted_image = sd15.lda.latents_to_image(x)
+
+    ensure_similar_images(predicted_image, expected_image_std_sde_random_init)
+
+
 @no_grad()
 def test_diffusion_batch2(sd15_std: StableDiffusion_1):
     sd15 = sd15_std

tests/e2e/test_diffusion_ref/README.md

@@ -67,6 +67,35 @@ Special cases:
 
 - `kitchen_dog.png` is generated with the same Diffusers script and negative prompt, seed 12, positive prompt "a small brown dog, detailed high-quality professional image, sitting on a chair, in a kitchen".
 
+- `expected_std_sde_random_init.png` is generated with the following code:
+
+```python
+import torch
+from diffusers import StableDiffusionPipeline
+from diffusers.schedulers.scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler
+
+from refiners.fluxion.utils import manual_seed
+
+diffusers_solver = DPMSolverMultistepScheduler.from_config(  # type: ignore
+    {
+        "beta_end": 0.012,
+        "beta_schedule": "scaled_linear",
+        "beta_start": 0.00085,
+        "algorithm_type": "sde-dpmsolver++",
+        "use_karras_sigmas": False,
+        "final_sigmas_type": "sigma_min",
+        "euler_at_final": True,
+    }
+)
+model_id = "runwayml/stable-diffusion-v1-5"
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32, scheduler=diffusers_solver)
+pipe = pipe.to("cuda")
+prompt = "a cute cat, detailed high-quality professional image"
+negative_prompt = "lowres, bad anatomy, bad hands, cropped, worst quality"
+manual_seed(2)
+image = pipe(prompt, negative_prompt=negative_prompt, guidance_scale=7.5).images[0]
+```
+
 - `kitchen_mask.png` is made manually.
 
 - Controlnet guides have been manually generated (x) using open source software and models, namely:

tests/foundationals/latent_diffusion/test_solvers.py

@@ -59,6 +59,45 @@ def test_dpm_solver_diffusers(n_steps: int, last_step_first_order: bool):
         assert allclose(diffusers_output, refiners_output, rtol=0.01), f"outputs differ at step {step}"
 
 
+@pytest.mark.parametrize("n_steps, last_step_first_order", [(5, False), (5, True), (30, False), (30, True)])
+def test_dpm_solver_sde_diffusers(n_steps: int, last_step_first_order: bool):
+    from diffusers import DPMSolverMultistepScheduler as DiffuserScheduler  # type: ignore
+
+    manual_seed(0)
+
+    diffusers_scheduler = DiffuserScheduler(
+        beta_schedule="scaled_linear",
+        beta_start=0.00085,
+        beta_end=0.012,
+        lower_order_final=False,
+        euler_at_final=last_step_first_order,
+        final_sigmas_type="sigma_min",  # default before Diffusers 0.26.0
+        algorithm_type="sde-dpmsolver++",
+    )
+    diffusers_scheduler.set_timesteps(n_steps)
+    solver = DPMSolver(
+        num_inference_steps=n_steps,
+        last_step_first_order=last_step_first_order,
+        params=SolverParams(sde_variance=1.0),
+    )
+    assert equal(solver.timesteps, diffusers_scheduler.timesteps)
+
+    sample = randn(1, 3, 32, 32)
+    predicted_noise = randn(1, 3, 32, 32)
+
+    manual_seed(37)
+    diffusers_outputs: list[Tensor] = [
+        cast(Tensor, diffusers_scheduler.step(predicted_noise, timestep, sample).prev_sample)  # type: ignore
+        for timestep in diffusers_scheduler.timesteps
+    ]
+
+    manual_seed(37)
+    refiners_outputs = [solver(x=sample, predicted_noise=predicted_noise, step=step) for step in range(n_steps)]
+
+    for step, (diffusers_output, refiners_output) in enumerate(zip(diffusers_outputs, refiners_outputs)):
+        assert allclose(diffusers_output, refiners_output, rtol=0.01, atol=1e-6), f"outputs differ at step {step}"
+
+
 def test_ddim_diffusers():
     from diffusers import DDIMScheduler  # type: ignore