Add an argument "negative_prompt"

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -116,6 +116,7 @@ def __call__(
         width: int = 512,
         num_inference_steps: int = 50,
         guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
         eta: float = 0.0,
         generator: Optional[torch.Generator] = None,
         latents: Optional[torch.FloatTensor] = None,
@@ -144,6 +145,8 @@ def __call__(
                 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation.
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
@@ -217,9 +220,25 @@ def __call__(
         do_classifier_free_guidance = guidance_scale > 1.0
         # get unconditional embeddings for classifier free guidance
         if do_classifier_free_guidance:
+            ucond_tokens: List[str]
+            if negative_prompt is None:
+                ucond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError("`negative_prompt` should be the same type to `prompt`.")
+            elif isinstance(negative_prompt, str):
+                ucond_tokens = [negative_prompt] * batch_size
+            elif batch_size != len(negative_prompt):
+                raise ValueError("The length of `negative_prompt` should be equal to batch_size.")
+            else:
+                ucond_tokens = negative_prompt
+
             max_length = text_input_ids.shape[-1]
             uncond_input = self.tokenizer(
-                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+                ucond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
             )
             uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -128,6 +128,7 @@ def __call__(
         strength: float = 0.8,
         num_inference_steps: Optional[int] = 50,
         guidance_scale: Optional[float] = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
         eta: Optional[float] = 0.0,
         generator: Optional[torch.Generator] = None,
         output_type: Optional[str] = "pil",
@@ -160,6 +161,8 @@ def __call__(
                 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation.
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
@@ -258,9 +261,25 @@ def __call__(
         do_classifier_free_guidance = guidance_scale > 1.0
         # get unconditional embeddings for classifier free guidance
         if do_classifier_free_guidance:
+            ucond_tokens: List[str]
+            if negative_prompt is None:
+                ucond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError("`negative_prompt` should be the same type to `prompt`.")
+            elif isinstance(negative_prompt, str):
+                ucond_tokens = [negative_prompt] * batch_size
+            elif batch_size != len(negative_prompt):
+                raise ValueError("The length of `negative_prompt` should be equal to batch_size.")
+            else:
+                ucond_tokens = negative_prompt
+
             max_length = text_input_ids.shape[-1]
             uncond_input = self.tokenizer(
-                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+                ucond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
             )
             uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -144,6 +144,7 @@ def __call__(
         strength: float = 0.8,
         num_inference_steps: Optional[int] = 50,
         guidance_scale: Optional[float] = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
         eta: Optional[float] = 0.0,
         generator: Optional[torch.Generator] = None,
         output_type: Optional[str] = "pil",
@@ -180,6 +181,8 @@ def __call__(
                 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                 usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation.
             eta (`float`, *optional*, defaults to 0.0):
                 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
                 [`schedulers.DDIMScheduler`], will be ignored for others.
@@ -292,9 +295,25 @@ def __call__(
         do_classifier_free_guidance = guidance_scale > 1.0
         # get unconditional embeddings for classifier free guidance
         if do_classifier_free_guidance:
+            ucond_tokens: List[str]
+            if negative_prompt is None:
+                ucond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError("`negative_prompt` should be the same type to `prompt`.")
+            elif isinstance(negative_prompt, str):
+                ucond_tokens = [negative_prompt] * batch_size
+            elif batch_size != len(negative_prompt):
+                raise ValueError("The length of `negative_prompt` should be equal to batch_size.")
+            else:
+                ucond_tokens = negative_prompt
+
             max_length = text_input_ids.shape[-1]
             uncond_input = self.tokenizer(
-                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
+                ucond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
             )
             uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
 

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_onnx.py

@@ -52,6 +52,7 @@ def __call__(
         width: Optional[int] = 512,
         num_inference_steps: Optional[int] = 50,
         guidance_scale: Optional[float] = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
         eta: Optional[float] = 0.0,
         latents: Optional[np.ndarray] = None,
         output_type: Optional[str] = "pil",
@@ -102,9 +103,25 @@ def __call__(
         do_classifier_free_guidance = guidance_scale > 1.0
         # get unconditional embeddings for classifier free guidance
         if do_classifier_free_guidance:
+            ucond_tokens: List[str]
+            if negative_prompt is None:
+                ucond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError("`negative_prompt` should be the same type to `prompt`.")
+            elif isinstance(negative_prompt, str):
+                ucond_tokens = [negative_prompt] * batch_size
+            elif batch_size != len(negative_prompt):
+                raise ValueError("The length of `negative_prompt` should be equal to batch_size.")
+            else:
+                ucond_tokens = negative_prompt
+
             max_length = text_input_ids.shape[-1]
             uncond_input = self.tokenizer(
-                [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np"
+                ucond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="np",
             )
             uncond_embeddings = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]