Add SVD (#5895)

* begin model

* finish blocks

* add_embedding

* addition_time_embed_dim

* use TimestepEmbedding

* fix temporal res block

* fix time_pos_embed

* fix add_embedding

* add conversion script

* fix model

* up

* add new resnet blocks

* make forward work

* return sample in original shape

* fix temb shape in TemporalResnetBlock

* add spatio temporal transformers

* add vae blocks

* fix blocks

* update

* update

* fix shapes in Alphablender and add time activation in res blcok

* use new blocks

* style

* fix temb shape

* fix SpatioTemporalResBlock

* reuse TemporalBasicTransformerBlock

* fix TemporalBasicTransformerBlock

* use TransformerSpatioTemporalModel

* fix TransformerSpatioTemporalModel

* fix time_context dim

* clean up

* make temb optional

* add blocks

* rename model

* update conversion script

* remove UNetMidBlockSpatioTemporal

* add in init

* remove unused arg

* remove unused arg

* remove more unsed args

* up

* up

* check for None

* update vae

* update up/mid blocks for decoder

* begin pipeline

* adapt scheduler

* add guidance scalings

* fix norm eps in temporal transformers

* add temporal autoencoder

* make pipeline run

* fix frame decodig

* decode in float32

* decode n frames at a time

* pass decoding_t to decode_latents

* fix decode_latents

* vae encode/decode in fp32

* fix dtype in TransformerSpatioTemporalModel

* type image_latents same as image_embeddings

* allow using differnt eps in temporal block for video decoder

* fix default values in vae

* pass num frames in decode

* switch spatial to temporal for mixing in VAE

* fix num frames during split decoding

* cast alpha to sample dtype

* fix attention in MidBlockTemporalDecoder

* fix typo

* fix guidance_scales dtype

* fix missing activation in TemporalDecoder

* skip_post_quant_conv

* add vae conversion

* style

* take guidance scale as input

* up

* allow passing PIL to export_video

* accept fps as arg

* add pipeline and vae in init

* remove hack

* use AutoencoderKLTemporalDecoder

* don't scale image latents

* add unet tests

* clean up unet

* clean TransformerSpatioTemporalModel

* add slow svd test

* clean up

* make temb optional in Decoder mid block

* fix norm eps in TransformerSpatioTemporalModel

* clean up temp decoder

* clean up

* clean up

* use c_noise values for timesteps

* use math for log

* update

* fix copies

* doc

* upcast vae

* update forward pass for gradient checkpointing

* make added_time_ids is tensor

* up

* fix upcasting

* remove post quant conv

* add _resize_with_antialiasing

* fix _compute_padding

* cleanup model

* more cleanup

* more cleanup

* more cleanup

* remove freeu

* remove attn slice

* small clean

* up

* up

* remove extra step kwargs

* remove eta

* remove dropout

* remove callback

* remove merge factor args

* clean

* clean up

* move to dedicated folder

* remove attention_head_dim

* docstr and small fix

* update unet doc strings

* rename decoding_t

* correct linting

* store c_skip and c_out

* cleanup

* clean TemporalResnetBlock

* more cleanup

* clean up vae

* clean up

* begin doc

* more cleanup

* up

* up

* doc

* Improve

* better naming

* better naming

* better naming

* better naming

* better naming

* better naming

* better naming

* better naming

* Apply suggestions from code review

* Default chunk size to None

* add example

* Better

* Apply suggestions from code review

* update doc

* Update src/diffusers/pipelines/stable_diffusion_video/pipeline_stable_diffusion_video.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* style

* Get torch compile working

* up

* rename

* fix doc

* add chunking

* torch compile

* torch compile

* add modelling outputs

* torch compile

* Improve chunking

* Apply suggestions from code review

* Update docs/source/en/using-diffusers/svd.md

* Close diff tag

* remove slicing

* resnet docstr

* add docstr in resnet

* rename

* Apply suggestions from code review

* update tests

* Fix output type latents

* fix more

* fix more

* Update docs/source/en/using-diffusers/svd.md

* fix more

* add pipeline tests

* remove unused arg

* clean  up

* make sure get_scaling receives tensors

* fix euler scheduler

* fix get_scalings

* simply euler for now

* remove old test file

* use randn_tensor to create noise

* fix device for rand tensor

* increase expected_max_difference

* fix test_inference_batch_single_identical

* actually fix test_inference_batch_single_identical

* disable test_save_load_float16

* skip test_float16_inference

* skip test_inference_batch_single_identical

* fix test_xformers_attention_forwardGenerator_pass

* Apply suggestions from code review

* update StableVideoDiffusionPipelineSlowTests

* update image

* add diffusers example

* fix more

---------
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: apolinário <joaopaulo.passos@gmail.com>

src/diffusers/schedulers/scheduling_deis_multistep.py

@@ -323,8 +323,20 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -358,8 +358,20 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

src/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py

@@ -358,8 +358,20 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

src/diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -357,8 +357,20 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

src/diffusers/schedulers/scheduling_euler_discrete.py

@@ -144,7 +144,10 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         interpolation_type: str = "linear",
         use_karras_sigmas: Optional[bool] = False,
+        sigma_min: Optional[float] = None,
+        sigma_max: Optional[float] = None,
         timestep_spacing: str = "linspace",
+        timestep_type: str = "discrete",  # can be "discrete" or "continuous"
         steps_offset: int = 0,
     ):
         if trained_betas is not None:
@@ -164,13 +167,22 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
 
         sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
-        sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32)
-        self.sigmas = torch.from_numpy(sigmas)
+        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
+
+        sigmas = torch.from_numpy(sigmas[::-1].copy()).to(dtype=torch.float32)
+        timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
 
         # setable values
         self.num_inference_steps = None
-        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
-        self.timesteps = torch.from_numpy(timesteps)
+
+        # TODO: Support the full EDM scalings for all prediction types and timestep types
+        if timestep_type == "continuous" and prediction_type == "v_prediction":
+            self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas])
+        else:
+            self.timesteps = timesteps
+
+        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+
         self.is_scale_input_called = False
         self.use_karras_sigmas = use_karras_sigmas
 
@@ -268,10 +280,15 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
 
-        sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
-        self.sigmas = torch.from_numpy(sigmas).to(device=device)
+        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
 
-        self.timesteps = torch.from_numpy(timesteps).to(device=device)
+        # TODO: Support the full EDM scalings for all prediction types and timestep types
+        if self.config.timestep_type == "continuous" and self.config.prediction_type == "v_prediction":
+            self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas]).to(device=device)
+        else:
+            self.timesteps = torch.from_numpy(timesteps.astype(np.float32)).to(device=device)
+
+        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
         self._step_index = None
 
     def _sigma_to_t(self, sigma, log_sigmas):
@@ -301,8 +318,20 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)
@@ -412,7 +441,7 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         elif self.config.prediction_type == "epsilon":
             pred_original_sample = sample - sigma_hat * model_output
         elif self.config.prediction_type == "v_prediction":
-            # * c_out + input * c_skip
+            # denoised = model_output * c_out + input * c_skip
             pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
         else:
             raise ValueError(

src/diffusers/schedulers/scheduling_heun_discrete.py

@@ -303,8 +303,20 @@ class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

src/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py

@@ -324,8 +324,20 @@ class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
     def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
-        sigma_min: float = in_sigmas[-1].item()
-        sigma_max: float = in_sigmas[0].item()
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
 
         rho = 7.0  # 7.0 is the value used in the paper
         ramp = np.linspace(0, 1, num_inference_steps)

tests/schedulers/test_scheduler_euler.py

@@ -37,6 +37,14 @@ class EulerDiscreteSchedulerTest(SchedulerCommonTest):
         for prediction_type in ["epsilon", "v_prediction"]:
             self.check_over_configs(prediction_type=prediction_type)
 
+    def test_timestep_type(self):
+        timestep_types = ["discrete", "continuous"]
+        for timestep_type in timestep_types:
+            self.check_over_configs(timestep_type=timestep_type)
+
+    def test_karras_sigmas(self):
+        self.check_over_configs(use_karras_sigmas=True, sigma_min=0.02, sigma_max=700.0)
+
     def test_full_loop_no_noise(self):
         scheduler_class = self.scheduler_classes[0]
         scheduler_config = self.get_scheduler_config()