rename to StableDiffusionT2IAdapterPipeline

Author: huo-ju

scripts/t2i_adapter_tester.py

@@ -1,38 +1,35 @@
 import cv2
 import argparse
-#import importlib
-import random
-import copy
 import torch
 import sys
 import os
 root_path = os.getcwd()
-print(root_path )
 sys.path.append(f"{root_path}/src")
 import diffusers
 from PIL import Image
-from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
+from diffusers import StableDiffusionT2IAdapterPipeline, DPMSolverMultistepScheduler
 from extra.t2iadapter.adapter import Adapter
 from basicsr.utils import img2tensor, tensor2img, scandir, get_time_str, get_root_logger, get_env_info
 
 diffusers.utils.logging.disable_progress_bar()
 
+
 class DummySafetyChecker():
     def safety_checker(self, images, *args, **kwargs):
         return images, [False] * len(images)
 
+
 def loadmodel(pipeline_name, model_path, **kwargs):
     print("load pipeline")
     print("load model from:", pipeline_name, model_path)
 
     kwargs["torch_dtype"] = torch.float16
-    if pipeline_name == "StableDiffusionPipeline":
-        kwargs["revision"] = "fp16"
+    kwargs["revision"] = "fp16"
 
     safechecker = DummySafetyChecker().safety_checker
     kwargs["safety_checker"] = safechecker
 
-    pipe = StableDiffusionPipeline.from_pretrained(model_path, **kwargs)
+    pipe = StableDiffusionT2IAdapterPipeline.from_pretrained(model_path, **kwargs)
     return pipe.to("cuda")
 
 
@@ -43,10 +40,8 @@ def generation(pipe, prompt, seed, features_adapter=None):
         "num_inference_steps": 50,
     }
     settings["prompt"] = prompt
-    g = torch.Generator(device="cuda")
-    settings["generator"] = g.manual_seed(seed)
     settings["features_adapter"] = features_adapter
-    settings["features_adapter_strength"] = 0.4
+    settings["features_adapter_strength"] = 0.5
     images = pipe(**settings).images
     return images
 
@@ -73,9 +68,7 @@ def main() -> int:
     features_adapter = model_ad(edge.to(device))
 
     pipe = loadmodel(args.pipeline, args.model_path, **kwargs)
-
-    if args.pipeline == "StableDiffusionPipeline":
-        pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
 
     prompt = f"A car with flying wings"
     outputimg = generation(pipe, prompt, 52, features_adapter)

src/diffusers/__init__.py

@@ -125,6 +125,7 @@
         StableDiffusionPipelineSafe,
         StableDiffusionPix2PixZeroPipeline,
         StableDiffusionSAGPipeline,
+        StableDiffusionT2IAdapterPipeline,
         StableDiffusionUpscalePipeline,
         StableUnCLIPImg2ImgPipeline,
         StableUnCLIPPipeline,

src/diffusers/pipelines/__init__.py

@@ -59,6 +59,7 @@
         StableDiffusionPipeline,
         StableDiffusionPix2PixZeroPipeline,
         StableDiffusionSAGPipeline,
+        StableDiffusionT2IAdapterPipeline,
         StableDiffusionUpscalePipeline,
         StableUnCLIPImg2ImgPipeline,
         StableUnCLIPPipeline,

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -50,6 +50,7 @@ class StableDiffusionPipelineOutput(BaseOutput):
     from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy
     from .pipeline_stable_diffusion_instruct_pix2pix import StableDiffusionInstructPix2PixPipeline
     from .pipeline_stable_diffusion_latent_upscale import StableDiffusionLatentUpscalePipeline
+    from .pipeline_stable_diffusion_t2iadapter import StableDiffusionT2IAdapterPipeline
     from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline
     from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
     from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -500,8 +500,6 @@ def __call__(
         eta: float = 0.0,
         generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
         latents: Optional[torch.FloatTensor] = None,
-        features_adapter: list = None,
-        features_adapter_strength: float = 0.4,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
         output_type: Optional[str] = "pil",
@@ -640,23 +638,13 @@ def __call__(
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-                # predict the noise residual
-
-                input_features_adapter = None
-                #num_inference_steps
-                if features_adapter_strength > 1:
-                    features_adapter_strength = 1
-                elif features_adapter_strength < 0:
-                    features_adapter_strength = 0
-                if i < int(num_inference_steps * features_adapter_strength):
-                    input_features_adapter = features_adapter
 
+                # predict the noise residual
                 noise_pred = self.unet(
                     latent_model_input,
                     t,
                     encoder_hidden_states=prompt_embeds,
                     cross_attention_kwargs=cross_attention_kwargs,
-                    features_adapter=input_features_adapter
                 ).sample
 
                 # perform guidance