[Bug fix] Make sure correct timesteps are chosen for img2img (#3128)

Make sure correct timesteps are chosen for img2img

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -503,7 +503,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start:]
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -528,7 +528,7 @@ class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start:]
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -390,7 +390,7 @@ class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoader
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start:]
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -511,7 +511,7 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMi
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start:]
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py

@@ -507,7 +507,7 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline, TextualInversionLo
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start:]
+        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 

tests/pipelines/stable_diffusion/test_stable_diffusion_img2img.py

@@ -25,6 +25,7 @@ from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
     DPMSolverMultistepScheduler,
+    HeunDiscreteScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     StableDiffusionImg2ImgPipeline,
@@ -416,6 +417,33 @@ class StableDiffusionImg2ImgPipelineSlowTests(unittest.TestCase):
         for module in pipe.text_encoder, pipe.unet, pipe.vae:
             assert module.device == torch.device("cpu")
 
+    def test_img2img_2nd_order(self):
+        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+        sd_pipe.scheduler = HeunDiscreteScheduler.from_config(sd_pipe.scheduler.config)
+        sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 10
+        inputs["strength"] = 0.75
+        image = sd_pipe(**inputs).images[0]
+
+        expected_image = load_numpy(
+            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/img2img_heun.npy"
+        )
+        max_diff = np.abs(expected_image - image).max()
+        assert max_diff < 5e-2
+
+        inputs = self.get_inputs(torch_device)
+        inputs["num_inference_steps"] = 11
+        inputs["strength"] = 0.75
+        image_other = sd_pipe(**inputs).images[0]
+
+        mean_diff = np.abs(image - image_other).mean()
+
+        # images should be very similar
+        assert mean_diff < 5e-2
+
     def test_stable_diffusion_img2img_pipeline_multiple_of_8(self):
         init_image = load_image(
             "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"