[LCM] Fix img2img (huggingface#5698)

* [LCM] Fix img2img

* make fix-copies

* make fix-copies

* make fix-copies

* up

Author: patrickvonplaten

src/diffusers/models/attention_processor.py

@@ -378,7 +378,7 @@ def set_processor(self, processor: "AttnProcessor", _remove_lora: bool = False)
             _remove_lora (`bool`, *optional*, defaults to `False`):
                 Set to `True` to remove LoRA layers from the model.
         """
-        if hasattr(self, "processor") and _remove_lora and self.to_q.lora_layer is not None:
+        if not USE_PEFT_BACKEND and hasattr(self, "processor") and _remove_lora and self.to_q.lora_layer is not None:
             deprecate(
                 "set_processor to offload LoRA",
                 "0.26.0",

src/diffusers/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py

@@ -738,7 +738,7 @@ def __call__(
             if original_inference_steps is not None
             else self.scheduler.config.original_inference_steps
         )
-        latent_timestep = torch.tensor(int(strength * original_inference_steps))
+        latent_timestep = timesteps[:1]
         latents = self.prepare_latents(
             image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
         )

tests/pipelines/latent_consistency_models/test_latent_consistency_models_img2img.py

@@ -133,7 +133,7 @@ def test_lcm_onestep(self):
         assert image.shape == (1, 32, 32, 3)
 
         image_slice = image[0, -3:, -3:, -1]
-        expected_slice = np.array([0.5865, 0.2854, 0.2828, 0.7473, 0.6006, 0.4580, 0.4397, 0.6415, 0.6069])
+        expected_slice = np.array([0.4388, 0.3717, 0.2202, 0.7213, 0.6370, 0.3664, 0.5815, 0.6080, 0.4977])
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
 
     def test_lcm_multistep(self):
@@ -150,7 +150,7 @@ def test_lcm_multistep(self):
         assert image.shape == (1, 32, 32, 3)
 
         image_slice = image[0, -3:, -3:, -1]
-        expected_slice = np.array([0.4903, 0.3304, 0.3503, 0.5241, 0.5153, 0.4585, 0.3222, 0.4764, 0.4891])
+        expected_slice = np.array([0.4150, 0.3719, 0.2479, 0.6333, 0.6024, 0.3778, 0.5036, 0.5420, 0.4678])
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
 
     def test_inference_batch_single_identical(self):
@@ -237,7 +237,7 @@ def test_lcm_onestep(self):
         assert image.shape == (1, 512, 512, 3)
 
         image_slice = image[0, -3:, -3:, -1].flatten()
-        expected_slice = np.array([0.1025, 0.0911, 0.0984, 0.0981, 0.0901, 0.0918, 0.1055, 0.0940, 0.0730])
+        expected_slice = np.array([0.1950, 0.1961, 0.2308, 0.1786, 0.1837, 0.2320, 0.1898, 0.1885, 0.2309])
         assert np.abs(image_slice - expected_slice).max() < 1e-3
 
     def test_lcm_multistep(self):
@@ -253,5 +253,5 @@ def test_lcm_multistep(self):
         assert image.shape == (1, 512, 512, 3)
 
         image_slice = image[0, -3:, -3:, -1].flatten()
-        expected_slice = np.array([0.01855, 0.01855, 0.01489, 0.01392, 0.01782, 0.01465, 0.01831, 0.02539, 0.0])
+        expected_slice = np.array([0.3756, 0.3816, 0.3767, 0.3718, 0.3739, 0.3735, 0.3863, 0.3803, 0.3563])
         assert np.abs(image_slice - expected_slice).max() < 1e-3