Add draft tests for DiffEdit

Author: clarencechen

tests/pipelines/stable_diffusion_2/test_stable_diffusion_diffedit.py

@@ -27,6 +27,8 @@
     AutoencoderKL,
     DDIMInverseScheduler,
     DDIMScheduler,
+    DPMSolverMultistepInverseScheduler,
+    DPMSolverMultistepScheduler,
     StableDiffusionDiffEditPipeline,
     UNet2DConditionModel,
 )
@@ -74,6 +76,13 @@ def get_dummy_components(self):
             clip_sample=False,
             set_alpha_to_zero=False,
         )
+        inverse_scheduler = DPMSolverMultistepInverseScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_zero=False,
+        )
         torch.manual_seed(0)
         vae = AutoencoderKL(
             block_out_channels=[32, 64],
@@ -249,6 +258,30 @@ def test_inversion(self):
         max_diff = np.abs(image_slice.flatten() - expected_slice).max()
         self.assertLessEqual(max_diff, 1e-3)
 
+    def test_inversion_dpm(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        components["scheduler"] = DPMSolverMultistepInverseScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+        )
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inversion_inputs(device)
+        image = pipe.invert(**inputs).images
+        image_slice = image[0, -1, -3:, -3:]
+
+        self.assertEqual(image.shape, (2, 32, 32, 3))
+        expected_slice = np.array(
+            [0.5150, 0.5134, 0.5043, 0.5376, 0.4694, 0.51050, 0.5015, 0.4407, 0.4799],
+        )
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        self.assertLessEqual(max_diff, 1e-3)
 
 @require_torch_gpu
 @slow
@@ -313,3 +346,54 @@ def test_stable_diffusion_diffedit_full(self):
             / 255
         )
         assert np.abs((expected_image - image).max()) < 5e-1
+
+    def test_stable_diffusion_diffedit_dpm(self):
+        generator = torch.manual_seed(0)
+
+        pipe = StableDiffusionDiffEditPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-2-1", safety_checker=None, torch_dtype=torch.float16
+        )
+        pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+        pipe.inverse_scheduler = DPMSolverMultistepInverseScheduler.from_config(pipe.scheduler.config)
+        pipe.enable_model_cpu_offload()
+        pipe.set_progress_bar_config(disable=None)
+
+        source_prompt = "a bowl of fruit"
+        target_prompt = "a bowl of pears"
+
+        mask_image = pipe.generate_mask(
+            image=self.raw_image,
+            source_prompt=source_prompt,
+            target_prompt=target_prompt,
+            generator=generator,
+        )
+
+        inv_latents = pipe.invert(
+            prompt=source_prompt,
+            image=self.raw_image,
+            inpaint_strength=0.7,
+            generator=generator,
+            num_inference_steps=25,
+        ).latents
+
+        image = pipe(
+            prompt=target_prompt,
+            mask_image=mask_image,
+            image_latents=inv_latents,
+            generator=generator,
+            negative_prompt=source_prompt,
+            inpaint_strength=0.7,
+            num_inference_steps=25,
+            output_type="numpy",
+        ).images[0]
+
+        expected_image = (
+            np.array(
+                load_image(
+                    "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+                    "/diffedit/pears.png"
+                ).resize((768, 768))
+            )
+            / 255
+        )
+        assert np.abs((expected_image - image).max()) < 5e-1