[PPDiffusers] StableDiffusion Pipeline 升级 (#390)

#377

完成1,2,3代码和单测，4和3单测相同
暂时测试可以，有其他问题再找时间修改

pytest test_stable_diffusion_img2img.py

![image](https://github.com/PaddlePaddle/PaddleMIX/assets/4617245/1e5bbbef-7090-43e4-b0ba-c2bfd13e98aa)

pytest test_stable_diffusion_inpaint.py

![image](https://github.com/PaddlePaddle/PaddleMIX/assets/4617245/7ca9d9b8-620d-4090-b09c-5b8653ab0d64)

pytest test_stable_diffusion_inpaint_legacy.py

![image](https://github.com/PaddlePaddle/PaddleMIX/assets/4617245/cd08e49f-6e90-4b7d-8c7a-6cffe7e22b90)

ppdiffusers/ppdiffusers/__init__.py

@@ -117,3 +117,8 @@
     from .utils.dummy_paddle_and_paddlenlp_and_einops_objects import *  # noqa F403
 else:
     from .pipelines import UniDiffuserPipeline
+    from .pipelines.stable_diffusion import (
+        StableDiffusionImg2ImgPipeline,
+        StableDiffusionInpaintPipeline,
+        StableDiffusionInpaintPipelineLegacy,
+    )

ppdiffusers/ppdiffusers/pipelines/__init__.py

@@ -28,7 +28,13 @@
 )
 from .lvdm import LVDMTextToVideoPipeline, LVDMUncondPipeline
 from .pipeline_utils import DiffusionPipeline
-from .stable_diffusion import StableDiffusionPipeline, StableDiffusionPipelineOutput
+from .stable_diffusion import (
+    StableDiffusionImg2ImgPipeline,
+    StableDiffusionInpaintPipeline,
+    StableDiffusionInpaintPipelineLegacy,
+    StableDiffusionPipeline,
+    StableDiffusionPipelineOutput,
+)
 from .stable_diffusion_xl import (
     StableDiffusionXLPipeline,
     StableDiffusionXLPipelineOutput,

ppdiffusers/ppdiffusers/pipelines/stable_diffusion/__init__.py

@@ -23,6 +23,11 @@
     StableDiffusionPipeline,
     StableDiffusionPipelineOutput,
 )
+from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
+from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline
+from .pipeline_stable_diffusion_inpaint_legacy import (
+    StableDiffusionInpaintPipelineLegacy,
+)
 from .safety_checker import StableDiffusionSafetyChecker
 
 try:

ppdiffusers/ppdiffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -389,7 +389,7 @@ def encode_prompt(
                 prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
 
         if self.text_encoder is not None:
-            prompt_embeds_dtype = self.text_encoder.dtype
+            prompt_embeds_dtype = self.text_encoder._dtype
         elif self.unet is not None:
             prompt_embeds_dtype = self.unet.dtype
         else:

ppdiffusers/ppdiffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py

@@ -276,8 +276,8 @@ def set_timesteps(
         self.sigmas_down = paddle.concat([sigmas_down[:1], sigmas_down[1:].repeat_interleave(2), sigmas_down[-1:]])
 
         timesteps = paddle.to_tensor(timesteps, dtype=paddle.float32)
-        sigmas_interpol = sigmas_interpol.cpu()
-        log_sigmas = self.log_sigmas.cpu()
+        sigmas_interpol = sigmas_interpol.cpu().numpy()
+        log_sigmas = self.log_sigmas.cpu().numpy()
         timesteps_interpol = np.array(
             [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol]
         )

ppdiffusers/ppdiffusers/schedulers/scheduling_k_dpm_2_discrete.py

@@ -268,8 +268,8 @@ def set_timesteps(
         timesteps = paddle.to_tensor(timesteps, dtype=paddle.float32)
 
         # interpolate timesteps
-        sigmas_interpol = sigmas_interpol.cpu()
-        log_sigmas = self.log_sigmas.cpu()
+        sigmas_interpol = sigmas_interpol.cpu().numpy()
+        log_sigmas = self.log_sigmas.cpu().numpy()
         timesteps_interpol = np.array(
             [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol]
         )

ppdiffusers/ppdiffusers/utils/__init__.py

@@ -160,6 +160,21 @@
     convert_state_dict_to_ppdiffusers,
     convert_unet_state_dict_to_peft,
 )
+from .testing_utils import (  # load_image,
+    floats_tensor,
+    image_grid,
+    load_hf_numpy,
+    load_numpy,
+    load_pd,
+    load_ppnlp_numpy,
+    nightly,
+    paddle_all_close,
+    paddle_device,
+    parse_flag_from_env,
+    print_tensor_test,
+    require_paddle_gpu,
+    slow,
+)
 
 logger = get_logger(__name__)
 

ppdiffusers/ppdiffusers/utils/paddle_utils.py

@@ -150,9 +150,7 @@ def randn_tensor(
 
         if isinstance(generator, (list, tuple)):
             batch_size = shape[0]
-            shape = [
-                1,
-            ] + shape[1:]
+            shape = (1,) + tuple(shape[1:])
             latents = [randn_pt(shape, generator=generator[i], dtype=dtype) for i in range(batch_size)]
             latents = paddle.concat(latents, axis=0)
         else:

ppdiffusers/tests/pipelines/stable_diffusion/__init__.py

@@ -0,0 +1,13 @@
+# Copyright (c) 2024 PaddlePaddle Authors. 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.