Merge branch 'latent-upscaler' of github.com:huggingface/diffusers into latent-upscaler

Currently implemented in EulerDiscreteScheduler. An alternative would
have been to initialize the scheduler with an array of `trained_betas`.
However, that is currently not possible because of #1367.

src/diffusers/__init__.py

@@ -73,6 +73,7 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipeline,
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
+        StableDiffusionUpscalerPipeline,
         VQDiffusionPipeline,
     )
 else:

src/diffusers/models/resnet.py

@@ -419,6 +419,8 @@ class ResnetBlock2D(nn.Module):
             self.nonlinearity = Mish()
         elif non_linearity == "silu":
             self.nonlinearity = nn.SiLU()
+        elif non_linearity == "gelu":
+            self.nonlinearity = nn.GELU()
 
         self.upsample = self.downsample = None
         if self.up:

src/diffusers/pipelines/__init__.py

@@ -23,6 +23,7 @@ if is_torch_available() and is_transformers_available():
         StableDiffusionInpaintPipeline,
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionPipeline,
+        StableDiffusionUpscalerPipeline,
     )
     from .vq_diffusion import VQDiffusionPipeline
 

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -33,6 +33,7 @@ if is_transformers_available() and is_torch_available():
     from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
     from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline
     from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy
+    from .pipeline_stable_diffusion_upscaler import StableDiffusionUpscalerPipeline
     from .safety_checker import StableDiffusionSafetyChecker
 
 if is_transformers_available() and is_onnx_available():

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -516,15 +516,20 @@ class StableDiffusionPipeline(DiffusionPipeline):
             if callback is not None and i % callback_steps == 0:
                 callback(i, t, latents)
 
-        # 8. Post-processing
-        image = self.decode_latents(latents)
+        if output_type == "latents":
+            # Skip safety checking if we are returning latents
+            image = latents
+            has_nsfw_concept = None
+        else:
+            # 8. Post-processing
+            image = self.decode_latents(latents)
 
-        # 9. Run safety checker
-        image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype)
+            # 9. Run safety checker
+            image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype)
 
-        # 10. Convert to PIL
-        if output_type == "pil":
-            image = self.numpy_to_pil(image)
+            # 10. Convert to PIL
+            if output_type == "pil":
+                image = self.numpy_to_pil(image)
 
         if not return_dict:
             return (image, has_nsfw_concept)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscaler.py

@@ -0,0 +1,443 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+from diffusers.utils import is_accelerate_available
+
+from ...configuration_utils import FrozenDict
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...pipeline_utils import DiffusionPipeline
+from ...schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from ...utils import deprecate, logging
+from . import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+# TODO: Remove when we migrate the upscaler model to diffusers >>>>>>>
+import k_diffusion as K
+import huggingface_hub
+
+from torch import nn
+import torch.nn.functional as F
+
+UPSCALER_REPO = "pcuenq/k-upscaler"
+
+class NoiseLevelAndTextConditionedUpscaler(nn.Module):
+    def __init__(self, inner_model, sigma_data=1., embed_dim=256):
+        super().__init__()
+        self.inner_model = inner_model
+        self.sigma_data = sigma_data
+        self.low_res_noise_embed = K.layers.FourierFeatures(1, embed_dim, std=2)
+
+    def forward(self, input, sigma, low_res, low_res_sigma, c, **kwargs):
+        cross_cond, cross_cond_padding, pooler = c
+        c_in = 1 / (low_res_sigma ** 2 + self.sigma_data ** 2) ** 0.5
+        c_noise = low_res_sigma.log1p()[:, None]
+        c_in = K.utils.append_dims(c_in, low_res.ndim)
+        low_res_noise_embed = self.low_res_noise_embed(c_noise)
+        low_res_in = F.interpolate(low_res, scale_factor=2, mode='nearest') * c_in
+        mapping_cond = torch.cat([low_res_noise_embed, pooler], dim=1)
+        return self.inner_model(input, sigma, unet_cond=low_res_in, mapping_cond=mapping_cond, cross_cond=cross_cond, cross_cond_padding=cross_cond_padding, **kwargs)
+
+def make_upscaler_model(config_path, model_path, pooler_dim=768, train=False, device='cpu'):
+    config = K.config.load_config(open(config_path))
+    model = K.config.make_model(config)
+    model = NoiseLevelAndTextConditionedUpscaler(
+        model,
+        sigma_data=config['model']['sigma_data'],
+        embed_dim=config['model']['mapping_cond_dim'] - pooler_dim,
+    )
+    ckpt = torch.load(model_path, map_location='cpu')
+    model.load_state_dict(ckpt['model_ema'])
+    model = K.config.make_denoiser_wrapper(config)(model)
+    if not train:
+        model = model.eval().requires_grad_(False)
+    return model.to(device)
+
+# <<<<<< To be removed when we migrate upscaler model to diffusers
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class StableDiffusionUpscalerPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with the upscaler model. Currently restricted to `EulerDiscreteScheduler`.nn
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        scheduler: EulerDiscreteScheduler,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPFeatureExtractor,
+    ):
+        super().__init__()
+        if safety_checker is None:
+            logger.warn(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+
+        # Download upscaler
+        config_path = huggingface_hub.hf_hub_download(UPSCALER_REPO, "config_laion_text_cond_latent_upscaler_2.json")
+        weights_path = huggingface_hub.hf_hub_download(UPSCALER_REPO, "laion_text_cond_latent_upscaler_2_1_00470000_slim.pth")
+        self.upscaler = make_upscaler_model(config_path, weights_path)
+
+    def to(self, torch_device: Optional[Union[str, torch.device]] = None):
+        self.upscaler.to(torch_device)
+        return super().to(torch_device)
+
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        # TODO: enable in the upscaler
+        pass
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        # TODO: disable in the upscaler
+        pass
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
+                `attention_head_dim` must be a multiple of `slice_size`.
+        """
+        pass
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def enable_sequential_cpu_offload(self):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device("cuda")
+
+        for cpu_offloaded_model in [self.text_encoder, self.vae, self.safety_checker, self.upscaler]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.vae.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def _get_text_conditioning(self, prompt, device, do_classifier_free_guidance):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        def get_conditioning(text):
+            text_inputs = self.tokenizer(
+                text,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids.to(device)
+            attention_mask = text_inputs.attention_mask.to(device)
+            cross_cond_padding = 1 - attention_mask
+
+            # I believe the attention mask should be provided here, but the original notebook does not do it
+            # TODO: test it out
+            # clip_output = self.text_encoder(input_ids=text_input_ids, attention_mask=attention_mask, output_hidden_states=True)
+            clip_output = self.text_encoder(input_ids=text_input_ids, output_hidden_states=True)
+            hidden_states = clip_output.hidden_states[-1]
+            pooler_output = clip_output.pooler_output
+
+            return hidden_states, cross_cond_padding.to(dtype=hidden_states.dtype), pooler_output
+
+        prompt_conditioning = get_conditioning(prompt)
+        conditioning = prompt_conditioning
+
+        if do_classifier_free_guidance:
+            uncond_conditioning = get_conditioning(batch_size * [""])
+            conditioning = [torch.cat([uc_item, c_item]) for uc_item, c_item in zip(uncond_conditioning, prompt_conditioning)]
+
+        return conditioning
+
+
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is not None:
+            safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        else:
+            has_nsfw_concept = None
+        return image, has_nsfw_concept
+
+    def decode_latents(self, latents):
+        latents = 1 / 0.18215 * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(self, latents, prompt):
+        batch, _, height, width = latents.shape
+        if height != width:
+            raise ValueError(f"Latents should be square, got {height}x{width} instead")
+        
+        if prompt is not None:
+            if not isinstance(prompt, str) and not isinstance(prompt, list):
+                raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+            if isinstance(prompt, list) and len(prompt) != batch:
+                raise ValueError(f"`prompt` length has to be equal to the latents batch_size ({batch}), but is {len(prompt)}")
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        latents: torch.FloatTensor,
+        prompt: Optional[Union[str, List[str]]] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        generator: Optional[torch.Generator] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            latents (`torch.FloatTensor`):
+                Latents to be upscaled. Generated from a Stable Diffusion Pipeline using `output_type="latents"`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image upscaling process.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            generator (`torch.Generator`, *optional*):
+                A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
+                deterministic.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+            When returning a tuple, the first element is a list with the generated images, and the second element is a
+            list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+            (nsfw) content, according to the `safety_checker`.
+        """
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(latents, prompt)
+
+        # 2. Define call parameters
+        batch_size, channels, height, width = latents.shape
+        if isinstance(prompt, str):
+            prompt = [prompt] * batch_size
+
+        device = self._execution_device
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Prepare timesteps
+        # We take log-linear steps in noise-level from sigma_max to sigma_min
+        # TODO(Pedro) Fix: create the scheduler with the betas instead
+        sigma_min = self.scheduler.sigmas[-2]   # Last one is zero
+        sigma_max = self.scheduler.sigmas[0]
+        # The +1 comes from k-diffusion
+        sigmas = torch.linspace(np.log(sigma_max), np.log(sigma_min), num_inference_steps+1).exp().to(device)
+        # scheduler.sigmas = torch.cat((sigmas, torch.tensor([0.]).to(device)))
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        self.scheduler.sigmas = sigmas
+
+        # # 4. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        # extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 5. Prepare conditioning
+        conditioning = self._get_text_conditioning(prompt, device, do_classifier_free_guidance)
+
+        # 6. Create initial noise
+        x_shape = [batch_size, channels, 2*height, 2*width]
+        noisy_latents = torch.randn(x_shape, generator=generator, device=device, dtype=sigmas.dtype)
+
+        # Disabled; according to the notebook it doesn't seem to work well
+        # TODO: remove in final implementation
+        low_res_sigma = torch.full([batch_size], 0, device=device, dtype=sigmas.dtype)
+
+        # 7. Prepare inputs for CFG
+        low_res = latents
+        if do_classifier_free_guidance:
+            low_res = torch.cat([low_res] * 2)
+            low_res_sigma = torch.cat([low_res_sigma] * 2)
+
+        # 8. Denoising loop
+        noisy_latents = noisy_latents * sigma_max
+        for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+            sigma = sigmas[i]
+            sigma = sigma[None]
+
+            latent_model_input = noisy_latents
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            if do_classifier_free_guidance:
+                latent_model_input = torch.cat([latent_model_input] * 2)
+                sigma = torch.cat([sigma] * 2)
+
+            # predict the next denoised latent
+            denoised = self.upscaler(
+                latent_model_input,
+                sigma,
+                low_res=low_res,
+                low_res_sigma=low_res_sigma,
+                c=conditioning,
+            )
+            
+            # perform guidance
+            if do_classifier_free_guidance:
+                uncond, cond = denoised.chunk(2)
+                denoised = uncond + guidance_scale * (cond - uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            noisy_latents = self.scheduler.step(denoised, t, noisy_latents).prev_sample
+
+        # 9. Post-processing
+        image = self.decode_latents(noisy_latents)
+
+        # 10. Run safety checker
+        image, has_nsfw_concept = self.run_safety_checker(image, device, sigmas.dtype)
+
+        # 11. Convert to PIL
+        if output_type == "pil":
+            image = self.numpy_to_pil(image)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

src/diffusers/schedulers/scheduling_euler_discrete.py

@@ -20,7 +20,7 @@ import torch
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..utils import _COMPATIBLE_STABLE_DIFFUSION_SCHEDULERS, BaseOutput, logging
-from .scheduling_utils import SchedulerMixin
+from .scheduling_utils import SchedulerMixin, betas_from_loglinear_sigmas
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
@@ -65,6 +65,8 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             `linear` or `scaled_linear`.
         trained_betas (`np.ndarray`, optional):
             option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+        predict_epsilon (`bool`):
+            optional flag to use when the model predicts the noise (epsilon), or the samples instead of the noise.
 
     """
 
@@ -78,6 +80,7 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         beta_end: float = 0.02,
         beta_schedule: str = "linear",
         trained_betas: Optional[np.ndarray] = None,
+        predict_epsilon: bool = True,
     ):
         if trained_betas is not None:
             self.betas = torch.from_numpy(trained_betas)
@@ -88,6 +91,11 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             self.betas = (
                 torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
             )
+        elif beta_schedule == "loglinear_sigmas":
+            # This scheduler is specific to k-diffusion latent upscaler
+            # We use a helper function because the computation is a bit involved
+            # Alternative: create from a list of `trained_betas` (but see https://github.com/huggingface/diffusers/issues/1367)
+            self.betas = betas_from_loglinear_sigmas(beta_start, beta_end, num_train_timesteps)
         else:
             raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
 
@@ -120,12 +128,15 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         Returns:
             `torch.FloatTensor`: scaled input sample
         """
+        self.is_scale_input_called = True
+        if not self.config.predict_epsilon:
+            return sample
+
         if isinstance(timestep, torch.Tensor):
             timestep = timestep.to(self.timesteps.device)
         step_index = (self.timesteps == timestep).nonzero().item()
         sigma = self.sigmas[step_index]
         sample = sample / ((sigma**2 + 1) ** 0.5)
-        self.is_scale_input_called = True
         return sample
 
     def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
@@ -229,7 +240,10 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
 
         # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
-        pred_original_sample = sample - sigma_hat * model_output
+        if self.config.predict_epsilon:
+            pred_original_sample = sample - sigma_hat * model_output
+        else:
+            pred_original_sample = model_output
 
         # 2. Convert to an ODE derivative
         derivative = (sample - pred_original_sample) / sigma_hat

src/diffusers/schedulers/scheduling_utils.py

@@ -16,6 +16,7 @@ import os
 from dataclasses import dataclass
 from typing import Any, Dict, Optional, Union
 
+import numpy as np
 import torch
 
 from ..utils import BaseOutput
@@ -152,3 +153,47 @@ class SchedulerMixin:
             getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c)
         ]
         return compatible_classes
+
+def betas_from_loglinear_sigmas(beta_start, beta_end, num_timesteps):
+    """
+    Computes the beta values suitable to create a loglinear schedule of sigmas,
+    as used in k-diffusion latent upscaler.
+
+    Concretely, these are the betas the create a sigma schedule like the following:
+    ```
+    torch.linspace(np.log(sigma_max), np.log(sigma_min), num_timesteps).exp()
+    ```
+
+    Args:
+        beta_start (`float`): The start sigma value.
+        beta_end (`float`): The end sigma value.
+        num_timesteps (`int`): The number of training timesteps.
+
+    Returns:
+        `torch.FloatTensor`: The beta values.
+    """
+    # First, compute sigma_max and sigma_min considering a "scaled_linear" schedule
+    # as used in Stable Diffusion. We just need sigma_min and sigma_max.
+    betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_timesteps) ** 2
+    alphas = 1 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
+    sigma_min, sigma_max = sigmas[0], sigmas[-1]
+
+    # Then, compute the actual loglinear sigmas from sigma_min, sigma_max
+    sigmas = torch.linspace(np.log(sigma_max), np.log(sigma_min), num_timesteps)
+    sigmas = sigmas.exp()
+    sigmas = np.array(sigmas)[::-1]
+    alpha_cumprod = 1./(1+sigmas**2)
+
+    # Compute the alpha values reversing alpha_cumprod
+    alphas = []
+    prev_prod = 1.
+    for a in alpha_cumprod:
+        current_alpha = a / prev_prod
+        alphas.append(current_alpha)
+        prev_prod = a
+    
+    # Get the betas from the alphas
+    betas = 1 - np.array(alphas)
+    return torch.Tensor(betas)