[WIP] Add img2img

src/diffusers/__init__.py

@@ -129,6 +129,7 @@
         IFPipeline,
         IFSuperResolutionPipeline,
         KandinskyInpaintPipeline,
+        KandinskyImg2ImgPipeline,
         KandinskyPipeline,
         KandinskyPriorPipeline,
         LDMTextToImagePipeline,

src/diffusers/pipelines/__init__.py

@@ -52,7 +52,7 @@
         IFPipeline,
         IFSuperResolutionPipeline,
     )
-    from .kandinsky import KandinskyInpaintPipeline, KandinskyPipeline, KandinskyPriorPipeline
+    from .kandinsky import KandinskyInpaintPipeline, KandinskyPipeline, KandinskyPriorPipeline, KandinskyImg2ImgPipeline
     from .latent_diffusion import LDMTextToImagePipeline
     from .paint_by_example import PaintByExamplePipeline
     from .semantic_stable_diffusion import SemanticStableDiffusionPipeline

src/diffusers/pipelines/kandinsky/__init__.py

@@ -14,6 +14,7 @@
 else:
     from .pipeline_kandinsky import KandinskyPipeline
     from .pipeline_kandinsky_inpaint import KandinskyInpaintPipeline
+    from .pipeline_kandinsky_img2img import KandinskyImg2ImgPipeline
     from .pipeline_kandinsky_prior import KandinskyPriorPipeline
     from .text_encoder import MultilingualCLIP
     from .text_proj import KandinskyTextProjModel

src/diffusers/pipelines/kandinsky/pipeline_kandinsky.py

@@ -22,7 +22,7 @@
 from ...models import UNet2DConditionModel, VQModel
 from ...pipelines import DiffusionPipeline
 from ...pipelines.pipeline_utils import ImagePipelineOutput
-from ...schedulers import UnCLIPScheduler
+from ...schedulers import DDPMScheduler
 from ...utils import (
     is_accelerate_available,
     is_accelerate_version,
@@ -58,7 +58,7 @@ class KandinskyPipeline(DiffusionPipeline):
             Frozen text-encoder.
         tokenizer ([`XLMRobertaTokenizer`]):
             Tokenizer of class
-        scheduler ([`UnCLIPScheduler`]):
+        scheduler ([`DDPMScheduler`]):
             A scheduler to be used in combination with `unet` to generate image latents.
         unet ([`UNet2DConditionModel`]):
             Conditional U-Net architecture to denoise the image embedding.
@@ -74,7 +74,7 @@ def __init__(
         tokenizer: XLMRobertaTokenizer,
         text_proj: KandinskyTextProjModel,
         unet: UNet2DConditionModel,
-        scheduler: UnCLIPScheduler,
+        scheduler: DDPMScheduler,
         movq: VQModel,
     ):
         super().__init__()
@@ -376,9 +376,7 @@ def __call__(
                 noise_pred,
                 t,
                 latent_model_input,
-                prev_timestep=prev_timestep,
                 generator=generator,
-                batch_size=batch_size,
             ).prev_sample
 
             _, latents = latents.chunk(2)