Remove unnecessary offset in img2img (#1653)

remove unnecessary offset in img2img

Author: patrickvonplaten

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -376,11 +376,9 @@ def check_inputs(self, prompt, strength, callback_steps):
 
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
-        offset = self.scheduler.config.get("steps_offset", 0)
-        init_timestep = int(num_inference_steps * strength) + offset
-        init_timestep = min(init_timestep, num_inference_steps)
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
-        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start:]
 
         return timesteps, num_inference_steps - t_start

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -414,11 +414,9 @@ def decode_latents(self, latents):
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
-        offset = self.scheduler.config.get("steps_offset", 0)
-        init_timestep = int(num_inference_steps * strength) + offset
-        init_timestep = min(init_timestep, num_inference_steps)
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
-        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start:]
 
         return timesteps, num_inference_steps - t_start

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py

@@ -323,11 +323,9 @@ def check_inputs(self, prompt, strength, callback_steps):
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
-        offset = self.scheduler.config.get("steps_offset", 0)
-        init_timestep = int(num_inference_steps * strength) + offset
-        init_timestep = min(init_timestep, num_inference_steps)
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
-        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start:]
 
         return timesteps, num_inference_steps - t_start

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -381,11 +381,9 @@ def check_inputs(self, prompt, strength, callback_steps):
 
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
-        offset = self.scheduler.config.get("steps_offset", 0)
-        init_timestep = int(num_inference_steps * strength) + offset
-        init_timestep = min(init_timestep, num_inference_steps)
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
-        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start:]
 
         return timesteps, num_inference_steps - t_start

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py

@@ -396,11 +396,9 @@ def check_inputs(self, prompt, strength, callback_steps):
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
     def get_timesteps(self, num_inference_steps, strength, device):
         # get the original timestep using init_timestep
-        offset = self.scheduler.config.get("steps_offset", 0)
-        init_timestep = int(num_inference_steps * strength) + offset
-        init_timestep = min(init_timestep, num_inference_steps)
+        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
-        t_start = max(num_inference_steps - init_timestep + offset, 0)
+        t_start = max(num_inference_steps - init_timestep, 0)
         timesteps = self.scheduler.timesteps[t_start:]
 
         return timesteps, num_inference_steps - t_start