feat(TAESD): support TAESD — Tiny Autoencoder for Stable Diffusion

invokeai/app/invocations/latent.py

@@ -1,11 +1,13 @@
 # Copyright (c) 2023 Kyle Schouviller (https://github.com/kyle0654)
 
 from contextlib import ExitStack
+from functools import singledispatchmethod
 from typing import List, Literal, Optional, Union
 
 import einops
 import torch
 import torchvision.transforms as T
+from diffusers import AutoencoderKL, AutoencoderTiny
 from diffusers.image_processor import VaeImageProcessor
 from diffusers.models.attention_processor import (
     AttnProcessor2_0,
@@ -15,7 +17,7 @@ from diffusers.models.attention_processor import (
 )
 from diffusers.schedulers import DPMSolverSDEScheduler
 from diffusers.schedulers import SchedulerMixin as Scheduler
-from pydantic import BaseModel, Field, validator
+from pydantic import validator
 from torchvision.transforms.functional import resize as tv_resize
 
 from invokeai.app.invocations.metadata import CoreMetadata
@@ -29,8 +31,21 @@ from invokeai.app.invocations.primitives import (
 from invokeai.app.util.controlnet_utils import prepare_control_image
 from invokeai.app.util.step_callback import stable_diffusion_step_callback
 from invokeai.backend.model_management.models import ModelType, SilenceWarnings
-
+from .baseinvocation import (
-from ...backend.model_management import BaseModelType, ModelPatcher
+    BaseInvocation,
+    FieldDescriptions,
+    Input,
+    InputField,
+    InvocationContext,
+    UIType,
+    tags,
+    title,
+)
+from .compel import ConditioningField
+from .controlnet_image_processors import ControlField
+from .model import ModelInfo, UNetField, VaeField
+from ..models.image import ImageCategory, ResourceOrigin
+from ...backend.model_management import BaseModelType
 from ...backend.model_management.lora import ModelPatcher
 from ...backend.stable_diffusion import PipelineIntermediateState
 from ...backend.stable_diffusion.diffusers_pipeline import (
@@ -42,22 +57,7 @@ from ...backend.stable_diffusion.diffusers_pipeline import (
 from ...backend.stable_diffusion.diffusion.shared_invokeai_diffusion import PostprocessingSettings
 from ...backend.stable_diffusion.schedulers import SCHEDULER_MAP
 from ...backend.util.devices import choose_precision, choose_torch_device
-from ..models.image import ImageCategory, ResourceOrigin
+from ...backend.util.logging import InvokeAILogger
-from .baseinvocation import (
-    BaseInvocation,
-    BaseInvocationOutput,
-    FieldDescriptions,
-    Input,
-    InputField,
-    InvocationContext,
-    OutputField,
-    UIType,
-    tags,
-    title,
-)
-from .compel import ConditioningField
-from .controlnet_image_processors import ControlField
-from .model import ModelInfo, UNetField, VaeField
 
 DEFAULT_PRECISION = choose_precision(choose_torch_device())
 
@@ -514,10 +514,17 @@ class LatentsToImageInvocation(BaseInvocation):
                 vae.to(dtype=torch.float16)
                 latents = latents.half()
 
+            try:
                 if self.tiled or context.services.configuration.tiled_decode:
                     vae.enable_tiling()
                 else:
                     vae.disable_tiling()
+            except AttributeError as err:
+                # FIXME: This is a TEMPORARY measure until AutoencoderTiny gets tiling support from https://github.com/huggingface/diffusers/pull/4627
+                if err.name.endswith("_tiling"):
+                    InvokeAILogger.getLogger(self.__class__.__name__).debug("ignoring tiling error for %s", vae.__class__, exc_info=err)
+                else:
+                    raise
 
             # clear memory as vae decode can request a lot
             torch.cuda.empty_cache()
@@ -704,16 +711,22 @@ class ImageToLatentsInvocation(BaseInvocation):
                 vae.to(dtype=torch.float16)
                 # latents = latents.half()
 
+            try:
                 if self.tiled:
                     vae.enable_tiling()
                 else:
                     vae.disable_tiling()
+            except AttributeError as err:
+                # FIXME: This is a TEMPORARY measure until AutoencoderTiny gets tiling support from https://github.com/huggingface/diffusers/pull/4627
+                if err.name.endswith("_tiling"):
+                    InvokeAILogger.getLogger(self.__class__.__name__).debug("ignoring tiling error for %s", vae.__class__, exc_info=err)
+                else:
+                    raise
 
             # non_noised_latents_from_image
             image_tensor = image_tensor.to(device=vae.device, dtype=vae.dtype)
             with torch.inference_mode():
-                image_tensor_dist = vae.encode(image_tensor).latent_dist
+                latents = self._encode_to_tensor(vae, image_tensor)
-                latents = image_tensor_dist.sample().to(dtype=vae.dtype)  # FIXME: uses torch.randn. make reproducible!
 
             latents = vae.config.scaling_factor * latents
             latents = latents.to(dtype=orig_dtype)
@@ -722,3 +735,12 @@ class ImageToLatentsInvocation(BaseInvocation):
         latents = latents.to("cpu")
         context.services.latents.save(name, latents)
         return build_latents_output(latents_name=name, latents=latents, seed=None)
+
+    def _encode_to_tensor(self, vae: AutoencoderKL, image_tensor: torch.FloatTensor) -> torch.FloatTensor:
+        image_tensor_dist = vae.encode(image_tensor).latent_dist
+        latents = image_tensor_dist.sample().to(dtype=vae.dtype)  # FIXME: uses torch.randn. make reproducible!
+        return latents
+
+    @singledispatchmethod
+    def _encode_to_tensor(self, vae: AutoencoderTiny, image_tensor: torch.FloatTensor) -> torch.FloatTensor:
+        return vae.encode(image_tensor).latents

invokeai/backend/model_management/model_probe.py

@@ -51,6 +51,7 @@ class ModelProbe(object):
         "StableDiffusionXLPipeline": ModelType.Main,
         "StableDiffusionXLImg2ImgPipeline": ModelType.Main,
         "AutoencoderKL": ModelType.Vae,
+        "AutoencoderTiny": ModelType.Vae,
         "ControlNetModel": ModelType.ControlNet,
     }